VictoriaMetrics/lib/storage/index_db.go

3515 lines
114 KiB
Go
Raw Normal View History

2019-05-22 21:16:55 +00:00
package storage
import (
"bytes"
"container/heap"
2019-05-22 21:16:55 +00:00
"errors"
"fmt"
"io"
"path/filepath"
"reflect"
2019-05-22 21:16:55 +00:00
"sort"
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 22:30:08 +00:00
"strconv"
2019-05-22 21:16:55 +00:00
"sync"
"sync/atomic"
"time"
"unsafe"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/encoding"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/fasttime"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/fs"
2019-05-22 21:16:55 +00:00
"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/memory"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/mergeset"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/querytracer"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/uint64set"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/workingsetcache"
2019-05-22 21:16:55 +00:00
"github.com/VictoriaMetrics/fastcache"
"github.com/cespare/xxhash/v2"
2019-05-22 21:16:55 +00:00
)
const (
// Prefix for MetricName->TSID entries.
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
//
// This index was substituted with nsPrefixDateMetricNameToTSID,
// since the MetricName->TSID index may require big amounts of memory for indexdb/dataBlocks cache
// when it grows big on the configured retention under high churn rate
// (e.g. when new time series are constantly registered).
//
// It is much more efficient from memory usage PoV to query per-day MetricName->TSID index
// (aka nsPrefixDateMetricNameToTSID) when the TSID must be obtained for the given MetricName
// during data ingestion under high churn rate and big retention.
//
// nsPrefixMetricNameToTSID = 0
2019-05-22 21:16:55 +00:00
// Prefix for Tag->MetricID entries.
nsPrefixTagToMetricIDs = 1
2019-05-22 21:16:55 +00:00
// Prefix for MetricID->TSID entries.
nsPrefixMetricIDToTSID = 2
// Prefix for MetricID->MetricName entries.
nsPrefixMetricIDToMetricName = 3
// Prefix for deleted MetricID entries.
nsPrefixDeletedMetricID = 4
2019-05-22 21:16:55 +00:00
// Prefix for Date->MetricID entries.
nsPrefixDateToMetricID = 5
// Prefix for (Date,Tag)->MetricID entries.
nsPrefixDateTagToMetricIDs = 6
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
// Prefix for (Date,MetricName)->TSID entries.
nsPrefixDateMetricNameToTSID = 7
2019-05-22 21:16:55 +00:00
)
// indexDB represents an index db.
type indexDB struct {
// Atomic counters must go at the top of the structure in order to properly align by 8 bytes on 32-bit archs.
// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/212 .
2019-05-22 21:16:55 +00:00
refCount uint64
// The number of missing MetricID -> TSID entries.
// High rate for this value means corrupted indexDB.
missingTSIDsForMetricID uint64
// The number of calls for date range searches.
dateRangeSearchCalls uint64
// The number of hits for date range searches.
dateRangeSearchHits uint64
// The number of calls for global search.
globalSearchCalls uint64
// missingMetricNamesForMetricID is a counter of missing MetricID -> MetricName entries.
// High rate may mean corrupted indexDB due to unclean shutdown.
// The db must be automatically recovered after that.
missingMetricNamesForMetricID uint64
mustDrop uint64
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 22:30:08 +00:00
// generation identifies the index generation ID
// and is used for syncing items from different indexDBs
generation uint64
name string
tb *mergeset.Table
2019-05-22 21:16:55 +00:00
extDB *indexDB
extDBLock sync.Mutex
// Cache for fast TagFilters -> MetricIDs lookup.
tagFiltersToMetricIDsCache *workingsetcache.Cache
2019-05-22 21:16:55 +00:00
// The parent storage.
s *Storage
// Cache for (date, tagFilter) -> loopsCount, which is used for reducing
// the amount of work when matching a set of filters.
loopsPerDateTagFilterCache *workingsetcache.Cache
2019-05-22 21:16:55 +00:00
indexSearchPool sync.Pool
}
var maxTagFiltersCacheSize int
// SetTagFiltersCacheSize overrides the default size of tagFiltersToMetricIDsCache
func SetTagFiltersCacheSize(size int) {
maxTagFiltersCacheSize = size
}
func getTagFiltersCacheSize() int {
if maxTagFiltersCacheSize <= 0 {
return int(float64(memory.Allowed()) / 32)
}
return maxTagFiltersCacheSize
}
// mustOpenIndexDB opens index db from the given path.
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 22:30:08 +00:00
//
// The last segment of the path should contain unique hex value which
// will be then used as indexDB.generation
lib/storage: pre-create timeseries before indexDB rotation (#4652) * lib/storage: pre-create timeseries before indexDB rotation during an hour before indexDB rotation start creating records at the next indexDB it must improve performance during switch for the next indexDB and remove ingestion issues. Since there is no need for creation new index records for timeseries already ingested into current indexDB https://github.com/VictoriaMetrics/VictoriaMetrics/issues/4563 * lib/storage: further work on indexdb rotation optimization - Document the change at docs/CHAGNELOG.md - Move back various caches from indexDB to Storage. This makes the change less intrusive. The dateMetricIDCache now takes into account indexDB generation, so it stores (date, metricID) entries for both the current and the next indexDB. - Consolidate the code responsible for idbNext pre-filling into prefillNextIndexDB() function. This improves code readability and maintainability a bit. - Rewrite and simplify the code responsible for calculating the next retention timestamp. Add various tests for corner cases of this code. - Remove indexdb pre-filling from RegisterMetricNames() function, since this function is rarely called. It is OK to add indexdb entries on demand in this function. This simplifies the code. Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 * docs/CHANGELOG.md: refer to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/4563 --------- Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2023-07-22 22:20:21 +00:00
func mustOpenIndexDB(path string, s *Storage, isReadOnly *uint32) *indexDB {
if s == nil {
logger.Panicf("BUG: Storage must be nin-nil")
}
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 22:30:08 +00:00
name := filepath.Base(path)
gen, err := strconv.ParseUint(name, 16, 64)
if err != nil {
logger.Panicf("FATAL: cannot parse indexdb path %q: %s", path, err)
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 22:30:08 +00:00
}
tb := mergeset.MustOpenTable(path, invalidateTagFiltersCache, mergeTagToMetricIDsRows, isReadOnly)
2019-05-22 21:16:55 +00:00
// Do not persist tagFiltersToMetricIDsCache in files, since it is very volatile.
2019-05-22 21:16:55 +00:00
mem := memory.Allowed()
tagFiltersCacheSize := getTagFiltersCacheSize()
2019-05-22 21:16:55 +00:00
db := &indexDB{
lib/storage: pre-create timeseries before indexDB rotation (#4652) * lib/storage: pre-create timeseries before indexDB rotation during an hour before indexDB rotation start creating records at the next indexDB it must improve performance during switch for the next indexDB and remove ingestion issues. Since there is no need for creation new index records for timeseries already ingested into current indexDB https://github.com/VictoriaMetrics/VictoriaMetrics/issues/4563 * lib/storage: further work on indexdb rotation optimization - Document the change at docs/CHAGNELOG.md - Move back various caches from indexDB to Storage. This makes the change less intrusive. The dateMetricIDCache now takes into account indexDB generation, so it stores (date, metricID) entries for both the current and the next indexDB. - Consolidate the code responsible for idbNext pre-filling into prefillNextIndexDB() function. This improves code readability and maintainability a bit. - Rewrite and simplify the code responsible for calculating the next retention timestamp. Add various tests for corner cases of this code. - Remove indexdb pre-filling from RegisterMetricNames() function, since this function is rarely called. It is OK to add indexdb entries on demand in this function. This simplifies the code. Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 * docs/CHANGELOG.md: refer to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/4563 --------- Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2023-07-22 22:20:21 +00:00
refCount: 1,
generation: gen,
tb: tb,
name: name,
2019-05-22 21:16:55 +00:00
tagFiltersToMetricIDsCache: workingsetcache.New(tagFiltersCacheSize),
s: s,
loopsPerDateTagFilterCache: workingsetcache.New(mem / 128),
2019-05-22 21:16:55 +00:00
}
return db
2019-05-22 21:16:55 +00:00
}
const noDeadline = 1<<64 - 1
2019-05-22 21:16:55 +00:00
// IndexDBMetrics contains essential metrics for indexDB.
type IndexDBMetrics struct {
TagFiltersToMetricIDsCacheSize uint64
TagFiltersToMetricIDsCacheSizeBytes uint64
TagFiltersToMetricIDsCacheSizeMaxBytes uint64
TagFiltersToMetricIDsCacheRequests uint64
TagFiltersToMetricIDsCacheMisses uint64
2019-05-22 21:16:55 +00:00
DeletedMetricsCount uint64
IndexDBRefCount uint64
MissingTSIDsForMetricID uint64
RecentHourMetricIDsSearchCalls uint64
RecentHourMetricIDsSearchHits uint64
DateRangeSearchCalls uint64
DateRangeSearchHits uint64
GlobalSearchCalls uint64
MissingMetricNamesForMetricID uint64
IndexBlocksWithMetricIDsProcessed uint64
IndexBlocksWithMetricIDsIncorrectOrder uint64
MinTimestampForCompositeIndex uint64
CompositeFilterSuccessConversions uint64
CompositeFilterMissingConversions uint64
2019-05-22 21:16:55 +00:00
mergeset.TableMetrics
}
func (db *indexDB) scheduleToDrop() {
atomic.AddUint64(&db.mustDrop, 1)
}
// UpdateMetrics updates m with metrics from the db.
func (db *indexDB) UpdateMetrics(m *IndexDBMetrics) {
var cs fastcache.Stats
cs.Reset()
db.tagFiltersToMetricIDsCache.UpdateStats(&cs)
m.TagFiltersToMetricIDsCacheSize += cs.EntriesCount
m.TagFiltersToMetricIDsCacheSizeBytes += cs.BytesSize
m.TagFiltersToMetricIDsCacheSizeMaxBytes += cs.MaxBytesSize
m.TagFiltersToMetricIDsCacheRequests += cs.GetCalls
m.TagFiltersToMetricIDsCacheMisses += cs.Misses
2019-05-22 21:16:55 +00:00
m.DeletedMetricsCount += uint64(db.s.getDeletedMetricIDs().Len())
2019-05-22 21:16:55 +00:00
m.IndexDBRefCount += atomic.LoadUint64(&db.refCount)
m.MissingTSIDsForMetricID += atomic.LoadUint64(&db.missingTSIDsForMetricID)
m.DateRangeSearchCalls += atomic.LoadUint64(&db.dateRangeSearchCalls)
m.DateRangeSearchHits += atomic.LoadUint64(&db.dateRangeSearchHits)
m.GlobalSearchCalls += atomic.LoadUint64(&db.globalSearchCalls)
m.MissingMetricNamesForMetricID += atomic.LoadUint64(&db.missingMetricNamesForMetricID)
m.IndexBlocksWithMetricIDsProcessed = atomic.LoadUint64(&indexBlocksWithMetricIDsProcessed)
m.IndexBlocksWithMetricIDsIncorrectOrder = atomic.LoadUint64(&indexBlocksWithMetricIDsIncorrectOrder)
m.MinTimestampForCompositeIndex = uint64(db.s.minTimestampForCompositeIndex)
m.CompositeFilterSuccessConversions = atomic.LoadUint64(&compositeFilterSuccessConversions)
m.CompositeFilterMissingConversions = atomic.LoadUint64(&compositeFilterMissingConversions)
2019-05-22 21:16:55 +00:00
db.tb.UpdateMetrics(&m.TableMetrics)
db.doExtDB(func(extDB *indexDB) {
extDB.tb.UpdateMetrics(&m.TableMetrics)
m.IndexDBRefCount += atomic.LoadUint64(&extDB.refCount)
})
}
func (db *indexDB) doExtDB(f func(extDB *indexDB)) bool {
db.extDBLock.Lock()
extDB := db.extDB
if extDB != nil {
extDB.incRef()
}
db.extDBLock.Unlock()
if extDB == nil {
return false
}
f(extDB)
extDB.decRef()
return true
}
// SetExtDB sets external db to search.
//
// It decrements refCount for the previous extDB.
func (db *indexDB) SetExtDB(extDB *indexDB) {
db.extDBLock.Lock()
prevExtDB := db.extDB
db.extDB = extDB
db.extDBLock.Unlock()
if prevExtDB != nil {
prevExtDB.decRef()
}
}
// MustClose closes db.
func (db *indexDB) MustClose() {
db.decRef()
}
func (db *indexDB) incRef() {
atomic.AddUint64(&db.refCount, 1)
}
func (db *indexDB) decRef() {
n := atomic.AddUint64(&db.refCount, ^uint64(0))
if int64(n) < 0 {
2019-05-22 21:16:55 +00:00
logger.Panicf("BUG: negative refCount: %d", n)
}
if n > 0 {
return
}
tbPath := db.tb.Path()
db.tb.MustClose()
db.SetExtDB(nil)
// Free space occupied by caches owned by db.
db.tagFiltersToMetricIDsCache.Stop()
db.loopsPerDateTagFilterCache.Stop()
db.tagFiltersToMetricIDsCache = nil
db.s = nil
db.loopsPerDateTagFilterCache = nil
2019-05-22 21:16:55 +00:00
if atomic.LoadUint64(&db.mustDrop) == 0 {
return
}
logger.Infof("dropping indexDB %q", tbPath)
fs.MustRemoveDirAtomic(tbPath)
2019-05-22 21:16:55 +00:00
logger.Infof("indexDB %q has been dropped", tbPath)
}
var tagBufPool bytesutil.ByteBufferPool
func (db *indexDB) getMetricIDsFromTagFiltersCache(qt *querytracer.Tracer, key []byte) ([]uint64, bool) {
qt = qt.NewChild("search for metricIDs in tag filters cache")
defer qt.Done()
buf := tagBufPool.Get()
defer tagBufPool.Put(buf)
buf.B = db.tagFiltersToMetricIDsCache.GetBig(buf.B[:0], key)
if len(buf.B) == 0 {
qt.Printf("cache miss")
return nil, false
}
qt.Printf("found metricIDs with size: %d bytes", len(buf.B))
metricIDs, err := unmarshalMetricIDs(nil, buf.B)
2019-05-22 21:16:55 +00:00
if err != nil {
logger.Panicf("FATAL: cannot unmarshal metricIDs from tagFiltersToMetricIDsCache: %s", err)
2019-05-22 21:16:55 +00:00
}
qt.Printf("unmarshaled %d metricIDs", len(metricIDs))
return metricIDs, true
2019-05-22 21:16:55 +00:00
}
func (db *indexDB) putMetricIDsToTagFiltersCache(qt *querytracer.Tracer, metricIDs []uint64, key []byte) {
qt = qt.NewChild("put %d metricIDs in cache", len(metricIDs))
defer qt.Done()
buf := tagBufPool.Get()
buf.B = marshalMetricIDs(buf.B, metricIDs)
qt.Printf("marshaled %d metricIDs into %d bytes", len(metricIDs), len(buf.B))
db.tagFiltersToMetricIDsCache.SetBig(key, buf.B)
qt.Printf("stored %d metricIDs into cache", len(metricIDs))
tagBufPool.Put(buf)
2019-05-22 21:16:55 +00:00
}
func (db *indexDB) getFromMetricIDCache(dst *TSID, metricID uint64) error {
2019-05-22 21:23:23 +00:00
// There is no need in prefixing the key with (accountID, projectID),
// since metricID is globally unique across all (accountID, projectID) values.
// See getUniqueUint64.
2019-05-22 21:16:55 +00:00
// There is no need in checking for deleted metricIDs here, since they
// must be checked by the caller.
buf := (*[unsafe.Sizeof(*dst)]byte)(unsafe.Pointer(dst))
key := (*[unsafe.Sizeof(metricID)]byte)(unsafe.Pointer(&metricID))
tmp := db.s.metricIDCache.Get(buf[:0], key[:])
2019-05-22 21:16:55 +00:00
if len(tmp) == 0 {
// The TSID for the given metricID wasn't found in the cache.
return io.EOF
}
if &tmp[0] != &buf[0] || len(tmp) != len(buf) {
return fmt.Errorf("corrupted MetricID->TSID cache: unexpected size for metricID=%d value; got %d bytes; want %d bytes", metricID, len(tmp), len(buf))
}
return nil
}
func (db *indexDB) putToMetricIDCache(metricID uint64, tsid *TSID) {
buf := (*[unsafe.Sizeof(*tsid)]byte)(unsafe.Pointer(tsid))
key := (*[unsafe.Sizeof(metricID)]byte)(unsafe.Pointer(&metricID))
db.s.metricIDCache.Set(key[:], buf[:])
2019-05-22 21:16:55 +00:00
}
func (db *indexDB) getMetricNameFromCache(dst []byte, metricID uint64) []byte {
2019-05-22 21:23:23 +00:00
// There is no need in prefixing the key with (accountID, projectID),
// since metricID is globally unique across all (accountID, projectID) values.
// See getUniqueUint64.
2019-05-22 21:16:55 +00:00
// There is no need in checking for deleted metricIDs here, since they
// must be checked by the caller.
key := (*[unsafe.Sizeof(metricID)]byte)(unsafe.Pointer(&metricID))
return db.s.metricNameCache.Get(dst, key[:])
2019-05-22 21:16:55 +00:00
}
func (db *indexDB) putMetricNameToCache(metricID uint64, metricName []byte) {
key := (*[unsafe.Sizeof(metricID)]byte)(unsafe.Pointer(&metricID))
db.s.metricNameCache.Set(key[:], metricName)
2019-05-22 21:16:55 +00:00
}
func marshalTagFiltersKey(dst []byte, tfss []*TagFilters, tr TimeRange, versioned bool) []byte {
prefix := ^uint64(0)
if versioned {
prefix = atomic.LoadUint64(&tagFiltersKeyGen)
2019-05-22 21:23:23 +00:00
}
// Round start and end times to per-day granularity according to per-day inverted index.
startDate := uint64(tr.MinTimestamp) / msecPerDay
endDate := uint64(tr.MaxTimestamp-1) / msecPerDay
2019-05-22 21:16:55 +00:00
dst = encoding.MarshalUint64(dst, prefix)
dst = encoding.MarshalUint64(dst, startDate)
dst = encoding.MarshalUint64(dst, endDate)
if len(tfss) == 0 {
return dst
}
dst = encoding.MarshalUint32(dst, tfss[0].accountID)
dst = encoding.MarshalUint32(dst, tfss[0].projectID)
2019-05-22 21:16:55 +00:00
for _, tfs := range tfss {
dst = append(dst, 0) // separator between tfs groups.
for i := range tfs.tfs {
2019-05-22 21:23:23 +00:00
dst = tfs.tfs[i].MarshalNoAccountIDProjectID(dst)
2019-05-22 21:16:55 +00:00
}
}
return dst
}
func invalidateTagFiltersCache() {
// This function must be fast, since it is called each
// time new timeseries is added.
atomic.AddUint64(&tagFiltersKeyGen, 1)
}
var tagFiltersKeyGen uint64
func marshalMetricIDs(dst []byte, metricIDs []uint64) []byte {
dst = encoding.MarshalUint64(dst, uint64(len(metricIDs)))
if len(metricIDs) == 0 {
return dst
2019-05-22 21:16:55 +00:00
}
var buf []byte
sh := (*reflect.SliceHeader)(unsafe.Pointer(&buf))
sh.Data = uintptr(unsafe.Pointer(&metricIDs[0]))
sh.Cap = 8 * len(metricIDs)
sh.Len = sh.Cap
dst = append(dst, buf...)
2019-05-22 21:16:55 +00:00
return dst
}
func unmarshalMetricIDs(dst []uint64, src []byte) ([]uint64, error) {
if len(src)%8 != 0 {
return dst, fmt.Errorf("cannot unmarshal metricIDs from buffer of %d bytes; the buffer length must divide by 8", len(src))
}
2019-05-22 21:16:55 +00:00
if len(src) < 8 {
return dst, fmt.Errorf("cannot unmarshal metricIDs len from buffer of %d bytes; need at least 8 bytes", len(src))
2019-05-22 21:16:55 +00:00
}
n := encoding.UnmarshalUint64(src)
if n > ((1<<64)-1)/8 {
return dst, fmt.Errorf("unexpectedly high metricIDs len: %d bytes; must be lower than %d bytes", n, uint64(((1<<64)-1)/8))
2019-05-22 21:16:55 +00:00
}
src = src[8:]
if n*8 != uint64(len(src)) {
return dst, fmt.Errorf("unexpected buffer length for unmarshaling metricIDs; got %d bytes; want %d bytes", n*8, len(src))
2019-05-22 21:16:55 +00:00
}
if n == 0 {
return dst, nil
2019-05-22 21:16:55 +00:00
}
var metricIDs []uint64
sh := (*reflect.SliceHeader)(unsafe.Pointer(&metricIDs))
sh.Data = uintptr(unsafe.Pointer(&src[0]))
sh.Cap = sh.Len
sh.Len = len(src) / 8
dst = append(dst, metricIDs...)
2019-05-22 21:16:55 +00:00
return dst, nil
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
// getTSIDByMetricName fills the dst with TSID for the given metricName at the given date.
2019-05-22 21:16:55 +00:00
//
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
// It returns false if the given metricName isn't found in the indexdb.
func (is *indexSearch) getTSIDByMetricName(dst *generationTSID, metricName []byte, date uint64) bool {
if is.getTSIDByMetricNameNoExtDB(&dst.TSID, metricName, date) {
// Fast path - the TSID is found in the current indexdb.
dst.generation = is.db.generation
return true
2019-05-22 21:16:55 +00:00
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
// Slow path - search for the TSID in the previous indexdb
ok := false
deadline := is.deadline
is.db.doExtDB(func(extDB *indexDB) {
is := extDB.getIndexSearch(0, 0, deadline)
ok = is.getTSIDByMetricNameNoExtDB(&dst.TSID, metricName, date)
extDB.putIndexSearch(is)
if ok {
dst.generation = extDB.generation
}
})
return ok
2019-05-22 21:16:55 +00:00
}
type indexSearch struct {
db *indexDB
ts mergeset.TableSearch
kb bytesutil.ByteBuffer
mp tagToMetricIDsRowParser
2019-05-22 21:16:55 +00:00
accountID uint32
projectID uint32
// deadline in unix timestamp seconds for the given search.
deadline uint64
2019-05-22 21:16:55 +00:00
}
func (db *indexDB) getIndexSearch(accountID, projectID uint32, deadline uint64) *indexSearch {
2019-05-22 21:16:55 +00:00
v := db.indexSearchPool.Get()
if v == nil {
v = &indexSearch{
db: db,
}
}
is := v.(*indexSearch)
is.ts.Init(db.tb)
is.accountID = accountID
is.projectID = projectID
is.deadline = deadline
2019-05-22 21:16:55 +00:00
return is
}
func (db *indexDB) putIndexSearch(is *indexSearch) {
is.ts.MustClose()
is.kb.Reset()
is.mp.Reset()
is.accountID = 0
is.projectID = 0
is.deadline = 0
2019-05-22 21:16:55 +00:00
db.indexSearchPool.Put(is)
}
lib/index: reduce read/write load after indexDB rotation (#2177) * lib/index: reduce read/write load after indexDB rotation IndexDB in VM is responsible for storing TSID - ID's used for identifying time series. The index is stored on disk and used by both ingestion and read path. IndexDB is stored separately to data parts and is global for all stored data. It can't be deleted partially as VM deletes data parts. Instead, indexDB is rotated once in `retention` interval. The rotation procedure means that `current` indexDB becomes `previous`, and new freshly created indexDB struct becomes `current`. So in any time, VM holds indexDB for current and previous retention periods. When time series is ingested or queried, VM checks if its TSID is present in `current` indexDB. If it is missing, it checks the `previous` indexDB. If TSID was found, it gets copied to the `current` indexDB. In this way `current` indexDB stores only series which were active during the retention period. To improve indexDB lookups, VM uses a cache layer called `tsidCache`. Both write and read path consult `tsidCache` and on miss the relad lookup happens. When rotation happens, VM resets the `tsidCache`. This is needed for ingestion path to trigger `current` indexDB re-population. Since index re-population requires additional resources, every index rotation event may cause some extra load on CPU and disk. While it may be unnoticeable for most of the cases, for systems with very high number of unique series each rotation may lead to performance degradation for some period of time. This PR makes an attempt to smooth out resource usage after the rotation. The changes are following: 1. `tsidCache` is no longer reset after the rotation; 2. Instead, each entry in `tsidCache` gains a notion of indexDB to which they belong; 3. On ingestion path after the rotation we check if requested TSID was found in `tsidCache`. Then we have 3 branches: 3.1 Fast path. It was found, and belongs to the `current` indexDB. Return TSID. 3.2 Slow path. It wasn't found, so we generate it from scratch, add to `current` indexDB, add it to `tsidCache`. 3.3 Smooth path. It was found but does not belong to the `current` indexDB. In this case, we add it to the `current` indexDB with some probability. The probability is based on time passed since the last rotation with some threshold. The more time has passed since rotation the higher is chance to re-populate `current` indexDB. The default re-population interval in this PR is set to `1h`, during which entries from `previous` index supposed to slowly re-populate `current` index. The new metric `vm_timeseries_repopulated_total` was added to identify how many TSIDs were moved from `previous` indexDB to the `current` indexDB. This metric supposed to grow only during the first `1h` after the last rotation. https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 Signed-off-by: hagen1778 <roman@victoriametrics.com> * wip * wip Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2022-02-11 22:30:08 +00:00
func generateTSID(dst *TSID, mn *MetricName) {
2019-05-22 21:23:23 +00:00
dst.AccountID = mn.AccountID
dst.ProjectID = mn.ProjectID
2019-05-22 21:16:55 +00:00
dst.MetricGroupID = xxhash.Sum64(mn.MetricGroup)
// Assume that the job-like metric is put at mn.Tags[0], while instance-like metric is put at mn.Tags[1]
// This assumption is true because mn.Tags must be sorted with mn.sortTags() before calling generateTSID() function.
// This allows grouping data blocks for the same (job, instance) close to each other on disk.
// This reduces disk seeks and disk read IO when data blocks are read from disk for the same job and/or instance.
2023-02-13 12:27:13 +00:00
// For example, data blocks for time series matching `process_resident_memory_bytes{job="vmstorage"}` are physically adjacent on disk.
2019-05-22 21:16:55 +00:00
if len(mn.Tags) > 0 {
dst.JobID = uint32(xxhash.Sum64(mn.Tags[0].Value))
}
if len(mn.Tags) > 1 {
dst.InstanceID = uint32(xxhash.Sum64(mn.Tags[1].Value))
}
dst.MetricID = generateUniqueMetricID()
2019-05-22 21:16:55 +00:00
}
func (is *indexSearch) createGlobalIndexes(tsid *TSID, mn *MetricName) {
2019-05-22 21:16:55 +00:00
// The order of index items is important.
// It guarantees index consistency.
ii := getIndexItems()
defer putIndexItems(ii)
2019-05-22 21:16:55 +00:00
// Create MetricID -> MetricName index.
ii.B = marshalCommonPrefix(ii.B, nsPrefixMetricIDToMetricName, mn.AccountID, mn.ProjectID)
ii.B = encoding.MarshalUint64(ii.B, tsid.MetricID)
ii.B = mn.Marshal(ii.B)
ii.Next()
2019-05-22 21:16:55 +00:00
// Create MetricID -> TSID index.
ii.B = marshalCommonPrefix(ii.B, nsPrefixMetricIDToTSID, mn.AccountID, mn.ProjectID)
ii.B = encoding.MarshalUint64(ii.B, tsid.MetricID)
ii.B = tsid.Marshal(ii.B)
ii.Next()
2019-05-22 21:16:55 +00:00
prefix := kbPool.Get()
prefix.B = marshalCommonPrefix(prefix.B[:0], nsPrefixTagToMetricIDs, mn.AccountID, mn.ProjectID)
ii.registerTagIndexes(prefix.B, mn, tsid.MetricID)
kbPool.Put(prefix)
is.db.tb.AddItems(ii.Items)
2019-05-22 21:16:55 +00:00
}
type indexItems struct {
B []byte
Items [][]byte
start int
}
func (ii *indexItems) reset() {
ii.B = ii.B[:0]
ii.Items = ii.Items[:0]
ii.start = 0
}
func (ii *indexItems) Next() {
ii.Items = append(ii.Items, ii.B[ii.start:])
ii.start = len(ii.B)
}
func getIndexItems() *indexItems {
v := indexItemsPool.Get()
if v == nil {
return &indexItems{}
}
return v.(*indexItems)
}
func putIndexItems(ii *indexItems) {
ii.reset()
indexItemsPool.Put(ii)
}
var indexItemsPool sync.Pool
// SearchLabelNamesWithFiltersOnTimeRange returns all the label names, which match the given tfss on the given tr.
func (db *indexDB) SearchLabelNamesWithFiltersOnTimeRange(qt *querytracer.Tracer, accountID, projectID uint32, tfss []*TagFilters, tr TimeRange,
maxLabelNames, maxMetrics int, deadline uint64) ([]string, error) {
qt = qt.NewChild("search for label names: filters=%s, timeRange=%s, maxLabelNames=%d, maxMetrics=%d", tfss, &tr, maxLabelNames, maxMetrics)
defer qt.Done()
lns := make(map[string]struct{})
qtChild := qt.NewChild("search for label names in the current indexdb")
is := db.getIndexSearch(accountID, projectID, deadline)
err := is.searchLabelNamesWithFiltersOnTimeRange(qtChild, lns, tfss, tr, maxLabelNames, maxMetrics)
db.putIndexSearch(is)
qtChild.Donef("found %d label names", len(lns))
if err != nil {
return nil, err
}
db.doExtDB(func(extDB *indexDB) {
qtChild := qt.NewChild("search for label names in the previous indexdb")
lnsLen := len(lns)
is := extDB.getIndexSearch(accountID, projectID, deadline)
err = is.searchLabelNamesWithFiltersOnTimeRange(qtChild, lns, tfss, tr, maxLabelNames, maxMetrics)
extDB.putIndexSearch(is)
qtChild.Donef("found %d additional label names", len(lns)-lnsLen)
})
if err != nil {
return nil, err
}
labelNames := make([]string, 0, len(lns))
for labelName := range lns {
labelNames = append(labelNames, labelName)
}
// Do not sort label names, since they must be sorted by vmselect.
qt.Printf("found %d label names in the current and the previous indexdb", len(labelNames))
return labelNames, nil
}
func (is *indexSearch) searchLabelNamesWithFiltersOnTimeRange(qt *querytracer.Tracer, lns map[string]struct{}, tfss []*TagFilters, tr TimeRange, maxLabelNames, maxMetrics int) error {
minDate := uint64(tr.MinTimestamp) / msecPerDay
maxDate := uint64(tr.MaxTimestamp-1) / msecPerDay
if maxDate == 0 || minDate > maxDate || maxDate-minDate > maxDaysForPerDaySearch {
qtChild := qt.NewChild("search for label names in global index: filters=%s", tfss)
err := is.searchLabelNamesWithFiltersOnDate(qtChild, lns, tfss, 0, maxLabelNames, maxMetrics)
qtChild.Done()
return err
}
var mu sync.Mutex
wg := getWaitGroup()
var errGlobal error
qt = qt.NewChild("parallel search for label names: filters=%s, timeRange=%s", tfss, &tr)
for date := minDate; date <= maxDate; date++ {
wg.Add(1)
qtChild := qt.NewChild("search for label names: filters=%s, date=%s", tfss, dateToString(date))
go func(date uint64) {
defer func() {
qtChild.Done()
wg.Done()
}()
lnsLocal := make(map[string]struct{})
isLocal := is.db.getIndexSearch(is.accountID, is.projectID, is.deadline)
err := isLocal.searchLabelNamesWithFiltersOnDate(qtChild, lnsLocal, tfss, date, maxLabelNames, maxMetrics)
is.db.putIndexSearch(isLocal)
mu.Lock()
defer mu.Unlock()
if errGlobal != nil {
return
}
if err != nil {
errGlobal = err
return
}
if len(lns) >= maxLabelNames {
return
}
for k := range lnsLocal {
lns[k] = struct{}{}
}
}(date)
}
wg.Wait()
putWaitGroup(wg)
qt.Done()
return errGlobal
}
func (is *indexSearch) searchLabelNamesWithFiltersOnDate(qt *querytracer.Tracer, lns map[string]struct{}, tfss []*TagFilters, date uint64, maxLabelNames, maxMetrics int) error {
filter, err := is.searchMetricIDsWithFiltersOnDate(qt, tfss, date, maxMetrics)
2019-05-22 21:16:55 +00:00
if err != nil {
return err
2019-05-22 21:16:55 +00:00
}
if filter != nil && filter.Len() <= 100e3 {
// It is faster to obtain label names by metricIDs from the filter
// instead of scanning the inverted index for the matching filters.
// This would help https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2978
metricIDs := filter.AppendTo(nil)
qt.Printf("sort %d metricIDs", len(metricIDs))
is.getLabelNamesForMetricIDs(qt, metricIDs, lns, maxLabelNames)
return nil
2019-05-22 21:16:55 +00:00
}
var prevLabelName []byte
2019-05-22 21:16:55 +00:00
ts := &is.ts
kb := &is.kb
mp := &is.mp
dmis := is.db.s.getDeletedMetricIDs()
loopsPaceLimiter := 0
underscoreNameSeen := false
nsPrefixExpected := byte(nsPrefixDateTagToMetricIDs)
if date == 0 {
nsPrefixExpected = nsPrefixTagToMetricIDs
}
kb.B = is.marshalCommonPrefixForDate(kb.B[:0], date)
prefix := kb.B
ts.Seek(prefix)
for len(lns) < maxLabelNames && ts.NextItem() {
if loopsPaceLimiter&paceLimiterFastIterationsMask == 0 {
if err := checkSearchDeadlineAndPace(is.deadline); err != nil {
return err
}
}
loopsPaceLimiter++
2019-05-22 21:16:55 +00:00
item := ts.Item
if !bytes.HasPrefix(item, prefix) {
2019-05-22 21:16:55 +00:00
break
}
if err := mp.Init(item, nsPrefixExpected); err != nil {
return err
2019-05-22 21:16:55 +00:00
}
if mp.GetMatchingSeriesCount(filter, dmis) == 0 {
continue
}
labelName := mp.Tag.Key
if len(labelName) == 0 {
underscoreNameSeen = true
}
if isArtificialTagKey(labelName) || string(labelName) == string(prevLabelName) {
// Search for the next tag key.
// The last char in kb.B must be tagSeparatorChar.
// Just increment it in order to jump to the next tag key.
kb.B = is.marshalCommonPrefixForDate(kb.B[:0], date)
if len(labelName) > 0 && labelName[0] == compositeTagKeyPrefix {
// skip composite tag entries
kb.B = append(kb.B, compositeTagKeyPrefix)
} else {
kb.B = marshalTagValue(kb.B, labelName)
}
kb.B[len(kb.B)-1]++
ts.Seek(kb.B)
continue
}
lns[string(labelName)] = struct{}{}
prevLabelName = append(prevLabelName[:0], labelName...)
2019-05-22 21:16:55 +00:00
}
if underscoreNameSeen {
lns["__name__"] = struct{}{}
}
2019-05-22 21:16:55 +00:00
if err := ts.Error(); err != nil {
return fmt.Errorf("error during search for prefix %q: %w", prefix, err)
2019-05-22 21:16:55 +00:00
}
return nil
}
func (is *indexSearch) getLabelNamesForMetricIDs(qt *querytracer.Tracer, metricIDs []uint64, lns map[string]struct{}, maxLabelNames int) {
lns["__name__"] = struct{}{}
var mn MetricName
foundLabelNames := 0
var buf []byte
for _, metricID := range metricIDs {
var ok bool
buf, ok = is.searchMetricNameWithCache(buf[:0], metricID)
if !ok {
// It is likely the metricID->metricName entry didn't propagate to inverted index yet.
// Skip this metricID for now.
continue
}
if err := mn.Unmarshal(buf); err != nil {
logger.Panicf("FATAL: cannot unmarshal metricName %q: %w", buf, err)
}
for _, tag := range mn.Tags {
if _, ok := lns[string(tag.Key)]; !ok {
foundLabelNames++
lns[string(tag.Key)] = struct{}{}
if len(lns) >= maxLabelNames {
qt.Printf("hit the limit on the number of unique label names: %d", maxLabelNames)
return
}
}
}
}
qt.Printf("get %d distinct label names from %d metricIDs", foundLabelNames, len(metricIDs))
}
// SearchTenants returns all tenants on the given tr.
func (db *indexDB) SearchTenants(qt *querytracer.Tracer, tr TimeRange, deadline uint64) ([]string, error) {
qt = qt.NewChild("search for tenants on timeRange=%s", &tr)
defer qt.Done()
tenants := make(map[string]struct{})
qtChild := qt.NewChild("search for tenants in the current indexdb")
is := db.getIndexSearch(0, 0, deadline)
err := is.searchTenantsOnTimeRange(qtChild, tenants, tr)
db.putIndexSearch(is)
qtChild.Donef("found %d tenants", len(tenants))
if err != nil {
return nil, err
}
db.doExtDB(func(extDB *indexDB) {
qtChild := qt.NewChild("search for tenants in the previous indexdb")
tenantsLen := len(tenants)
is := extDB.getIndexSearch(0, 0, deadline)
err = is.searchTenantsOnTimeRange(qtChild, tenants, tr)
extDB.putIndexSearch(is)
qtChild.Donef("found %d additional tenants", len(tenants)-tenantsLen)
})
if err != nil {
return nil, err
}
tenantsList := make([]string, 0, len(tenants))
for tenant := range tenants {
tenantsList = append(tenantsList, tenant)
}
// Do not sort tenants, since they must be sorted by vmselect.
qt.Printf("found %d tenants in the current and the previous indexdb", len(tenantsList))
return tenantsList, nil
}
func (is *indexSearch) searchTenantsOnTimeRange(qt *querytracer.Tracer, tenants map[string]struct{}, tr TimeRange) error {
minDate := uint64(tr.MinTimestamp) / msecPerDay
maxDate := uint64(tr.MaxTimestamp-1) / msecPerDay
if maxDate == 0 || minDate > maxDate || maxDate-minDate > maxDaysForPerDaySearch {
qtChild := qt.NewChild("search for tenants in global index")
err := is.searchTenantsOnDate(tenants, 0)
qtChild.Done()
return err
}
var mu sync.Mutex
wg := getWaitGroup()
var errGlobal error
qt = qt.NewChild("parallel search for tenants on timeRange=%s", &tr)
for date := minDate; date <= maxDate; date++ {
wg.Add(1)
qtChild := qt.NewChild("search for tenants on date=%s", dateToString(date))
go func(date uint64) {
defer func() {
qtChild.Done()
wg.Done()
}()
tenantsLocal := make(map[string]struct{})
isLocal := is.db.getIndexSearch(0, 0, is.deadline)
err := isLocal.searchTenantsOnDate(tenantsLocal, date)
is.db.putIndexSearch(isLocal)
mu.Lock()
defer mu.Unlock()
if errGlobal != nil {
return
}
if err != nil {
errGlobal = err
return
}
for tenant := range tenantsLocal {
tenants[tenant] = struct{}{}
}
}(date)
}
wg.Wait()
putWaitGroup(wg)
qt.Done()
return errGlobal
}
func (is *indexSearch) searchTenantsOnDate(tenants map[string]struct{}, date uint64) error {
loopsPaceLimiter := 0
ts := &is.ts
kb := &is.kb
is.accountID = 0
is.projectID = 0
kb.B = is.marshalCommonPrefixForDate(kb.B[:0], date)
_, prefixNeeded, _, _, err := unmarshalCommonPrefix(kb.B)
if err != nil {
logger.Panicf("BUG: cannot unmarshal common prefix from %q: %s", kb.B, err)
}
ts.Seek(kb.B)
for ts.NextItem() {
if loopsPaceLimiter&paceLimiterFastIterationsMask == 0 {
if err := checkSearchDeadlineAndPace(is.deadline); err != nil {
return err
}
}
loopsPaceLimiter++
_, prefix, accountID, projectID, err := unmarshalCommonPrefix(ts.Item)
if err != nil {
return err
}
if prefix != prefixNeeded {
// Reached the end of enteris with the needed prefix.
break
}
tenant := fmt.Sprintf("%d:%d", accountID, projectID)
tenants[tenant] = struct{}{}
// Seek for the next (accountID, projectID)
projectID++
if projectID == 0 {
accountID++
if accountID == 0 {
// Reached the end (accountID, projectID) space
break
}
}
is.accountID = accountID
is.projectID = projectID
kb.B = is.marshalCommonPrefixForDate(kb.B[:0], date)
ts.Seek(kb.B)
}
if err := ts.Error(); err != nil {
return fmt.Errorf("error during search for prefix %q: %w", kb.B, err)
}
return nil
}
// SearchLabelValuesWithFiltersOnTimeRange returns label values for the given labelName, tfss and tr.
func (db *indexDB) SearchLabelValuesWithFiltersOnTimeRange(qt *querytracer.Tracer, accountID, projectID uint32, labelName string, tfss []*TagFilters, tr TimeRange,
maxLabelValues, maxMetrics int, deadline uint64) ([]string, error) {
qt = qt.NewChild("search for label values: labelName=%q, filters=%s, timeRange=%s, maxLabelNames=%d, maxMetrics=%d", labelName, tfss, &tr, maxLabelValues, maxMetrics)
defer qt.Done()
lvs := make(map[string]struct{})
qtChild := qt.NewChild("search for label values in the current indexdb")
is := db.getIndexSearch(accountID, projectID, deadline)
err := is.searchLabelValuesWithFiltersOnTimeRange(qtChild, lvs, labelName, tfss, tr, maxLabelValues, maxMetrics)
db.putIndexSearch(is)
qtChild.Donef("found %d label values", len(lvs))
if err != nil {
return nil, err
}
db.doExtDB(func(extDB *indexDB) {
qtChild := qt.NewChild("search for label values in the previous indexdb")
lvsLen := len(lvs)
is := extDB.getIndexSearch(accountID, projectID, deadline)
err = is.searchLabelValuesWithFiltersOnTimeRange(qtChild, lvs, labelName, tfss, tr, maxLabelValues, maxMetrics)
extDB.putIndexSearch(is)
qtChild.Donef("found %d additional label values", len(lvs)-lvsLen)
})
if err != nil {
return nil, err
}
labelValues := make([]string, 0, len(lvs))
for labelValue := range lvs {
if len(labelValue) == 0 {
// Skip empty values, since they have no any meaning.
// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/600
continue
}
labelValues = append(labelValues, labelValue)
}
// Do not sort labelValues, since they must be sorted by vmselect.
qt.Printf("found %d label values in the current and the previous indexdb", len(labelValues))
return labelValues, nil
}
func (is *indexSearch) searchLabelValuesWithFiltersOnTimeRange(qt *querytracer.Tracer, lvs map[string]struct{}, labelName string, tfss []*TagFilters,
tr TimeRange, maxLabelValues, maxMetrics int) error {
minDate := uint64(tr.MinTimestamp) / msecPerDay
maxDate := uint64(tr.MaxTimestamp-1) / msecPerDay
if maxDate == 0 || minDate > maxDate || maxDate-minDate > maxDaysForPerDaySearch {
qtChild := qt.NewChild("search for label values in global index: labelName=%q, filters=%s", labelName, tfss)
err := is.searchLabelValuesWithFiltersOnDate(qtChild, lvs, labelName, tfss, 0, maxLabelValues, maxMetrics)
qtChild.Done()
return err
}
var mu sync.Mutex
wg := getWaitGroup()
var errGlobal error
qt = qt.NewChild("parallel search for label values: labelName=%q, filters=%s, timeRange=%s", labelName, tfss, &tr)
for date := minDate; date <= maxDate; date++ {
wg.Add(1)
qtChild := qt.NewChild("search for label names: filters=%s, date=%s", tfss, dateToString(date))
go func(date uint64) {
defer func() {
qtChild.Done()
wg.Done()
}()
lvsLocal := make(map[string]struct{})
isLocal := is.db.getIndexSearch(is.accountID, is.projectID, is.deadline)
err := isLocal.searchLabelValuesWithFiltersOnDate(qtChild, lvsLocal, labelName, tfss, date, maxLabelValues, maxMetrics)
is.db.putIndexSearch(isLocal)
mu.Lock()
defer mu.Unlock()
if errGlobal != nil {
return
}
if err != nil {
errGlobal = err
return
}
if len(lvs) >= maxLabelValues {
return
}
for v := range lvsLocal {
lvs[v] = struct{}{}
}
}(date)
}
wg.Wait()
putWaitGroup(wg)
qt.Done()
return errGlobal
}
func (is *indexSearch) searchLabelValuesWithFiltersOnDate(qt *querytracer.Tracer, lvs map[string]struct{}, labelName string, tfss []*TagFilters,
date uint64, maxLabelValues, maxMetrics int) error {
filter, err := is.searchMetricIDsWithFiltersOnDate(qt, tfss, date, maxMetrics)
2019-05-22 21:16:55 +00:00
if err != nil {
return err
2019-05-22 21:16:55 +00:00
}
if filter != nil && filter.Len() < 100e3 {
// It is faster to obtain label values by metricIDs from the filter
// instead of scanning the inverted index for the matching filters.
// This would help https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2978
metricIDs := filter.AppendTo(nil)
qt.Printf("sort %d metricIDs", len(metricIDs))
is.getLabelValuesForMetricIDs(qt, lvs, labelName, metricIDs, maxLabelValues)
return nil
2019-05-22 21:16:55 +00:00
}
if labelName == "__name__" {
// __name__ label is encoded as empty string in indexdb.
labelName = ""
2019-05-22 21:16:55 +00:00
}
labelNameBytes := bytesutil.ToUnsafeBytes(labelName)
var prevLabelValue []byte
2019-05-22 21:16:55 +00:00
ts := &is.ts
kb := &is.kb
mp := &is.mp
dmis := is.db.s.getDeletedMetricIDs()
loopsPaceLimiter := 0
nsPrefixExpected := byte(nsPrefixDateTagToMetricIDs)
if date == 0 {
nsPrefixExpected = nsPrefixTagToMetricIDs
}
kb.B = is.marshalCommonPrefixForDate(kb.B[:0], date)
kb.B = marshalTagValue(kb.B, labelNameBytes)
prefix := kb.B
2019-05-22 21:16:55 +00:00
ts.Seek(prefix)
for len(lvs) < maxLabelValues && ts.NextItem() {
if loopsPaceLimiter&paceLimiterFastIterationsMask == 0 {
if err := checkSearchDeadlineAndPace(is.deadline); err != nil {
return err
}
}
loopsPaceLimiter++
item := ts.Item
if !bytes.HasPrefix(item, prefix) {
2019-05-22 21:16:55 +00:00
break
}
if err := mp.Init(item, nsPrefixExpected); err != nil {
return err
2019-05-22 21:16:55 +00:00
}
if mp.GetMatchingSeriesCount(filter, dmis) == 0 {
continue
}
labelValue := mp.Tag.Value
if string(labelValue) == string(prevLabelValue) {
// Search for the next tag value.
// The last char in kb.B must be tagSeparatorChar.
// Just increment it in order to jump to the next tag value.
kb.B = is.marshalCommonPrefixForDate(kb.B[:0], date)
kb.B = marshalTagValue(kb.B, labelNameBytes)
kb.B = marshalTagValue(kb.B, labelValue)
kb.B[len(kb.B)-1]++
ts.Seek(kb.B)
continue
}
lvs[string(labelValue)] = struct{}{}
prevLabelValue = append(prevLabelValue[:0], labelValue...)
2019-05-22 21:16:55 +00:00
}
if err := ts.Error(); err != nil {
return fmt.Errorf("error when searching for tag name prefix %q: %w", prefix, err)
2019-05-22 21:16:55 +00:00
}
return nil
}
func (is *indexSearch) getLabelValuesForMetricIDs(qt *querytracer.Tracer, lvs map[string]struct{}, labelName string, metricIDs []uint64, maxLabelValues int) {
if labelName == "" {
labelName = "__name__"
}
var mn MetricName
foundLabelValues := 0
var buf []byte
for _, metricID := range metricIDs {
var ok bool
buf, ok = is.searchMetricNameWithCache(buf[:0], metricID)
if !ok {
// It is likely the metricID->metricName entry didn't propagate to inverted index yet.
// Skip this metricID for now.
continue
}
if err := mn.Unmarshal(buf); err != nil {
logger.Panicf("FATAL: cannot unmarshal metricName %q: %s", buf, err)
}
tagValue := mn.GetTagValue(labelName)
if _, ok := lvs[string(tagValue)]; !ok {
foundLabelValues++
lvs[string(tagValue)] = struct{}{}
if len(lvs) >= maxLabelValues {
qt.Printf("hit the limit on the number of unique label values for label %q: %d", labelName, maxLabelValues)
return
}
}
}
qt.Printf("get %d distinct values for label %q from %d metricIDs", foundLabelValues, labelName, len(metricIDs))
}
// SearchTagValueSuffixes returns all the tag value suffixes for the given tagKey and tagValuePrefix on the given tr.
//
// This allows implementing https://graphite-api.readthedocs.io/en/latest/api.html#metrics-find or similar APIs.
//
// If it returns maxTagValueSuffixes suffixes, then it is likely more than maxTagValueSuffixes suffixes is found.
func (db *indexDB) SearchTagValueSuffixes(qt *querytracer.Tracer, accountID, projectID uint32, tr TimeRange, tagKey, tagValuePrefix string,
delimiter byte, maxTagValueSuffixes int, deadline uint64) ([]string, error) {
qt = qt.NewChild("search tag value suffixes for accountID=%d, projectID=%d, timeRange=%s, tagKey=%q, tagValuePrefix=%q, delimiter=%c, maxTagValueSuffixes=%d",
accountID, projectID, &tr, tagKey, tagValuePrefix, delimiter, maxTagValueSuffixes)
defer qt.Done()
// TODO: cache results?
tvss := make(map[string]struct{})
is := db.getIndexSearch(accountID, projectID, deadline)
err := is.searchTagValueSuffixesForTimeRange(tvss, tr, tagKey, tagValuePrefix, delimiter, maxTagValueSuffixes)
db.putIndexSearch(is)
if err != nil {
return nil, err
}
if len(tvss) < maxTagValueSuffixes {
db.doExtDB(func(extDB *indexDB) {
is := extDB.getIndexSearch(accountID, projectID, deadline)
qtChild := qt.NewChild("search tag value suffixes in the previous indexdb")
err = is.searchTagValueSuffixesForTimeRange(tvss, tr, tagKey, tagValuePrefix, delimiter, maxTagValueSuffixes)
qtChild.Done()
extDB.putIndexSearch(is)
})
if err != nil {
return nil, err
}
}
suffixes := make([]string, 0, len(tvss))
for suffix := range tvss {
// Do not skip empty suffixes, since they may represent leaf tag values.
suffixes = append(suffixes, suffix)
}
if len(suffixes) > maxTagValueSuffixes {
suffixes = suffixes[:maxTagValueSuffixes]
}
// Do not sort suffixes, since they must be sorted by vmselect.
qt.Printf("found %d suffixes", len(suffixes))
return suffixes, nil
}
func (is *indexSearch) searchTagValueSuffixesForTimeRange(tvss map[string]struct{}, tr TimeRange, tagKey, tagValuePrefix string, delimiter byte, maxTagValueSuffixes int) error {
minDate := uint64(tr.MinTimestamp) / msecPerDay
maxDate := uint64(tr.MaxTimestamp-1) / msecPerDay
if minDate > maxDate || maxDate-minDate > maxDaysForPerDaySearch {
return is.searchTagValueSuffixesAll(tvss, tagKey, tagValuePrefix, delimiter, maxTagValueSuffixes)
}
// Query over multiple days in parallel.
wg := getWaitGroup()
var errGlobal error
var mu sync.Mutex // protects tvss + errGlobal from concurrent access below.
for minDate <= maxDate {
wg.Add(1)
go func(date uint64) {
defer wg.Done()
tvssLocal := make(map[string]struct{})
isLocal := is.db.getIndexSearch(is.accountID, is.projectID, is.deadline)
err := isLocal.searchTagValueSuffixesForDate(tvssLocal, date, tagKey, tagValuePrefix, delimiter, maxTagValueSuffixes)
is.db.putIndexSearch(isLocal)
mu.Lock()
defer mu.Unlock()
if errGlobal != nil {
return
}
if err != nil {
errGlobal = err
return
}
if len(tvss) > maxTagValueSuffixes {
return
}
for k := range tvssLocal {
tvss[k] = struct{}{}
}
}(minDate)
minDate++
}
wg.Wait()
putWaitGroup(wg)
return errGlobal
}
func (is *indexSearch) searchTagValueSuffixesAll(tvss map[string]struct{}, tagKey, tagValuePrefix string, delimiter byte, maxTagValueSuffixes int) error {
kb := &is.kb
nsPrefix := byte(nsPrefixTagToMetricIDs)
kb.B = is.marshalCommonPrefix(kb.B[:0], nsPrefix)
kb.B = marshalTagValue(kb.B, bytesutil.ToUnsafeBytes(tagKey))
kb.B = marshalTagValue(kb.B, bytesutil.ToUnsafeBytes(tagValuePrefix))
kb.B = kb.B[:len(kb.B)-1] // remove tagSeparatorChar from the end of kb.B
prefix := append([]byte(nil), kb.B...)
return is.searchTagValueSuffixesForPrefix(tvss, nsPrefix, prefix, len(tagValuePrefix), delimiter, maxTagValueSuffixes)
}
func (is *indexSearch) searchTagValueSuffixesForDate(tvss map[string]struct{}, date uint64, tagKey, tagValuePrefix string, delimiter byte, maxTagValueSuffixes int) error {
nsPrefix := byte(nsPrefixDateTagToMetricIDs)
kb := &is.kb
kb.B = is.marshalCommonPrefix(kb.B[:0], nsPrefix)
kb.B = encoding.MarshalUint64(kb.B, date)
kb.B = marshalTagValue(kb.B, bytesutil.ToUnsafeBytes(tagKey))
kb.B = marshalTagValue(kb.B, bytesutil.ToUnsafeBytes(tagValuePrefix))
kb.B = kb.B[:len(kb.B)-1] // remove tagSeparatorChar from the end of kb.B
prefix := append([]byte(nil), kb.B...)
return is.searchTagValueSuffixesForPrefix(tvss, nsPrefix, prefix, len(tagValuePrefix), delimiter, maxTagValueSuffixes)
}
func (is *indexSearch) searchTagValueSuffixesForPrefix(tvss map[string]struct{}, nsPrefix byte, prefix []byte, tagValuePrefixLen int, delimiter byte, maxTagValueSuffixes int) error {
kb := &is.kb
ts := &is.ts
mp := &is.mp
dmis := is.db.s.getDeletedMetricIDs()
loopsPaceLimiter := 0
ts.Seek(prefix)
for len(tvss) < maxTagValueSuffixes && ts.NextItem() {
if loopsPaceLimiter&paceLimiterFastIterationsMask == 0 {
if err := checkSearchDeadlineAndPace(is.deadline); err != nil {
return err
}
}
loopsPaceLimiter++
item := ts.Item
if !bytes.HasPrefix(item, prefix) {
break
}
if err := mp.Init(item, nsPrefix); err != nil {
return err
}
if mp.GetMatchingSeriesCount(nil, dmis) == 0 {
continue
}
tagValue := mp.Tag.Value
suffix := tagValue[tagValuePrefixLen:]
n := bytes.IndexByte(suffix, delimiter)
if n < 0 {
// Found leaf tag value that doesn't have delimiters after the given tagValuePrefix.
tvss[string(suffix)] = struct{}{}
continue
}
// Found non-leaf tag value. Extract suffix that end with the given delimiter.
suffix = suffix[:n+1]
tvss[string(suffix)] = struct{}{}
if suffix[len(suffix)-1] == 255 {
continue
}
// Search for the next suffix
suffix[len(suffix)-1]++
kb.B = append(kb.B[:0], prefix...)
kb.B = marshalTagValue(kb.B, suffix)
kb.B = kb.B[:len(kb.B)-1] // remove tagSeparatorChar
ts.Seek(kb.B)
}
if err := ts.Error(); err != nil {
return fmt.Errorf("error when searching for tag value sufixes for prefix %q: %w", prefix, err)
}
return nil
}
2019-05-22 21:23:23 +00:00
// GetSeriesCount returns the approximate number of unique timeseries for the given (accountID, projectID).
2019-05-22 21:16:55 +00:00
//
// It includes the deleted series too and may count the same series
// up to two times - in db and extDB.
func (db *indexDB) GetSeriesCount(accountID, projectID uint32, deadline uint64) (uint64, error) {
is := db.getIndexSearch(accountID, projectID, deadline)
n, err := is.getSeriesCount()
2019-05-22 21:16:55 +00:00
db.putIndexSearch(is)
if err != nil {
return 0, err
}
var nExt uint64
db.doExtDB(func(extDB *indexDB) {
is := extDB.getIndexSearch(accountID, projectID, deadline)
nExt, err = is.getSeriesCount()
2019-05-22 21:16:55 +00:00
extDB.putIndexSearch(is)
})
if err != nil {
return 0, fmt.Errorf("error when searching in extDB: %w", err)
2019-05-22 21:16:55 +00:00
}
return n + nExt, nil
}
func (is *indexSearch) getSeriesCount() (uint64, error) {
ts := &is.ts
kb := &is.kb
mp := &is.mp
loopsPaceLimiter := 0
var metricIDsLen uint64
// Extract the number of series from ((__name__=value): metricIDs) rows
kb.B = is.marshalCommonPrefix(kb.B[:0], nsPrefixTagToMetricIDs)
kb.B = marshalTagValue(kb.B, nil)
ts.Seek(kb.B)
for ts.NextItem() {
if loopsPaceLimiter&paceLimiterFastIterationsMask == 0 {
if err := checkSearchDeadlineAndPace(is.deadline); err != nil {
return 0, err
}
}
loopsPaceLimiter++
item := ts.Item
if !bytes.HasPrefix(item, kb.B) {
break
}
tail := item[len(kb.B):]
n := bytes.IndexByte(tail, tagSeparatorChar)
if n < 0 {
return 0, fmt.Errorf("invalid tag->metricIDs line %q: cannot find tagSeparatorChar %d", item, tagSeparatorChar)
}
tail = tail[n+1:]
if err := mp.InitOnlyTail(item, tail); err != nil {
return 0, err
}
// Take into account deleted timeseries too.
// It is OK if series can be counted multiple times in rare cases -
// the returned number is an estimation.
metricIDsLen += uint64(mp.MetricIDsLen())
}
if err := ts.Error(); err != nil {
return 0, fmt.Errorf("error when counting unique timeseries: %w", err)
}
return metricIDsLen, nil
}
// GetTSDBStatus returns topN entries for tsdb status for the given tfss, date and focusLabel.
func (db *indexDB) GetTSDBStatus(qt *querytracer.Tracer, accountID, projectID uint32, tfss []*TagFilters, date uint64, focusLabel string, topN, maxMetrics int, deadline uint64) (*TSDBStatus, error) {
qtChild := qt.NewChild("collect tsdb stats in the current indexdb")
is := db.getIndexSearch(accountID, projectID, deadline)
status, err := is.getTSDBStatus(qtChild, tfss, date, focusLabel, topN, maxMetrics)
qtChild.Done()
db.putIndexSearch(is)
if err != nil {
return nil, err
}
if status.hasEntries() {
return status, nil
}
db.doExtDB(func(extDB *indexDB) {
qtChild := qt.NewChild("collect tsdb stats in the previous indexdb")
is := extDB.getIndexSearch(accountID, projectID, deadline)
status, err = is.getTSDBStatus(qtChild, tfss, date, focusLabel, topN, maxMetrics)
qtChild.Done()
extDB.putIndexSearch(is)
})
if err != nil {
return nil, fmt.Errorf("error when obtaining TSDB status from extDB: %w", err)
}
return status, nil
}
// getTSDBStatus returns topN entries for tsdb status for the given tfss, date and focusLabel.
func (is *indexSearch) getTSDBStatus(qt *querytracer.Tracer, tfss []*TagFilters, date uint64, focusLabel string, topN, maxMetrics int) (*TSDBStatus, error) {
filter, err := is.searchMetricIDsWithFiltersOnDate(qt, tfss, date, maxMetrics)
if err != nil {
return nil, err
}
if filter != nil && filter.Len() == 0 {
qt.Printf("no matching series for filter=%s", tfss)
return &TSDBStatus{}, nil
}
ts := &is.ts
kb := &is.kb
mp := &is.mp
dmis := is.db.s.getDeletedMetricIDs()
thSeriesCountByMetricName := newTopHeap(topN)
thSeriesCountByLabelName := newTopHeap(topN)
thSeriesCountByFocusLabelValue := newTopHeap(topN)
thSeriesCountByLabelValuePair := newTopHeap(topN)
thLabelValueCountByLabelName := newTopHeap(topN)
var tmp, prevLabelName, prevLabelValuePair []byte
var labelValueCountByLabelName, seriesCountByLabelValuePair uint64
var totalSeries, labelSeries, totalLabelValuePairs uint64
nameEqualBytes := []byte("__name__=")
focusLabelEqualBytes := []byte(focusLabel + "=")
loopsPaceLimiter := 0
nsPrefixExpected := byte(nsPrefixDateTagToMetricIDs)
if date == 0 {
nsPrefixExpected = nsPrefixTagToMetricIDs
}
kb.B = is.marshalCommonPrefixForDate(kb.B[:0], date)
prefix := kb.B
ts.Seek(prefix)
for ts.NextItem() {
if loopsPaceLimiter&paceLimiterFastIterationsMask == 0 {
if err := checkSearchDeadlineAndPace(is.deadline); err != nil {
return nil, err
}
}
loopsPaceLimiter++
item := ts.Item
if !bytes.HasPrefix(item, prefix) {
break
}
if err := mp.Init(item, nsPrefixExpected); err != nil {
return nil, err
}
matchingSeriesCount := mp.GetMatchingSeriesCount(filter, dmis)
if matchingSeriesCount == 0 {
// Skip rows without matching metricIDs.
continue
}
tmp = append(tmp[:0], mp.Tag.Key...)
labelName := tmp
if isArtificialTagKey(labelName) {
// Skip artificially created tag keys.
kb.B = append(kb.B[:0], prefix...)
if len(labelName) > 0 && labelName[0] == compositeTagKeyPrefix {
kb.B = append(kb.B, compositeTagKeyPrefix)
} else {
kb.B = marshalTagValue(kb.B, labelName)
}
kb.B[len(kb.B)-1]++
ts.Seek(kb.B)
continue
}
if len(labelName) == 0 {
labelName = append(labelName, "__name__"...)
tmp = labelName
}
if string(labelName) == "__name__" {
totalSeries += uint64(matchingSeriesCount)
}
tmp = append(tmp, '=')
tmp = append(tmp, mp.Tag.Value...)
labelValuePair := tmp
if len(prevLabelName) == 0 {
prevLabelName = append(prevLabelName[:0], labelName...)
}
if string(labelName) != string(prevLabelName) {
thLabelValueCountByLabelName.push(prevLabelName, labelValueCountByLabelName)
thSeriesCountByLabelName.push(prevLabelName, labelSeries)
labelSeries = 0
labelValueCountByLabelName = 0
prevLabelName = append(prevLabelName[:0], labelName...)
}
if len(prevLabelValuePair) == 0 {
prevLabelValuePair = append(prevLabelValuePair[:0], labelValuePair...)
labelValueCountByLabelName++
}
if string(labelValuePair) != string(prevLabelValuePair) {
thSeriesCountByLabelValuePair.push(prevLabelValuePair, seriesCountByLabelValuePair)
if bytes.HasPrefix(prevLabelValuePair, nameEqualBytes) {
thSeriesCountByMetricName.push(prevLabelValuePair[len(nameEqualBytes):], seriesCountByLabelValuePair)
}
if bytes.HasPrefix(prevLabelValuePair, focusLabelEqualBytes) {
thSeriesCountByFocusLabelValue.push(prevLabelValuePair[len(focusLabelEqualBytes):], seriesCountByLabelValuePair)
}
seriesCountByLabelValuePair = 0
labelValueCountByLabelName++
prevLabelValuePair = append(prevLabelValuePair[:0], labelValuePair...)
}
// It is OK if series can be counted multiple times in rare cases -
// the returned number is an estimation.
labelSeries += uint64(matchingSeriesCount)
seriesCountByLabelValuePair += uint64(matchingSeriesCount)
totalLabelValuePairs += uint64(matchingSeriesCount)
}
if err := ts.Error(); err != nil {
return nil, fmt.Errorf("error when counting time series by metric names: %w", err)
}
thLabelValueCountByLabelName.push(prevLabelName, labelValueCountByLabelName)
thSeriesCountByLabelName.push(prevLabelName, labelSeries)
thSeriesCountByLabelValuePair.push(prevLabelValuePair, seriesCountByLabelValuePair)
if bytes.HasPrefix(prevLabelValuePair, nameEqualBytes) {
thSeriesCountByMetricName.push(prevLabelValuePair[len(nameEqualBytes):], seriesCountByLabelValuePair)
}
if bytes.HasPrefix(prevLabelValuePair, focusLabelEqualBytes) {
thSeriesCountByFocusLabelValue.push(prevLabelValuePair[len(focusLabelEqualBytes):], seriesCountByLabelValuePair)
}
status := &TSDBStatus{
TotalSeries: totalSeries,
TotalLabelValuePairs: totalLabelValuePairs,
SeriesCountByMetricName: thSeriesCountByMetricName.getSortedResult(),
SeriesCountByLabelName: thSeriesCountByLabelName.getSortedResult(),
SeriesCountByFocusLabelValue: thSeriesCountByFocusLabelValue.getSortedResult(),
SeriesCountByLabelValuePair: thSeriesCountByLabelValuePair.getSortedResult(),
LabelValueCountByLabelName: thLabelValueCountByLabelName.getSortedResult(),
}
return status, nil
}
// TSDBStatus contains TSDB status data for /api/v1/status/tsdb.
//
// See https://prometheus.io/docs/prometheus/latest/querying/api/#tsdb-stats
type TSDBStatus struct {
TotalSeries uint64
TotalLabelValuePairs uint64
SeriesCountByMetricName []TopHeapEntry
SeriesCountByLabelName []TopHeapEntry
SeriesCountByFocusLabelValue []TopHeapEntry
SeriesCountByLabelValuePair []TopHeapEntry
LabelValueCountByLabelName []TopHeapEntry
}
func (status *TSDBStatus) hasEntries() bool {
return len(status.SeriesCountByLabelValuePair) > 0
}
// topHeap maintains a heap of topHeapEntries with the maximum TopHeapEntry.n values.
type topHeap struct {
topN int
a []TopHeapEntry
}
// newTopHeap returns topHeap for topN items.
func newTopHeap(topN int) *topHeap {
return &topHeap{
topN: topN,
}
}
// TopHeapEntry represents an entry from `top heap` used in stats.
type TopHeapEntry struct {
Name string
Count uint64
}
func (th *topHeap) push(name []byte, count uint64) {
if count == 0 {
return
}
if len(th.a) < th.topN {
th.a = append(th.a, TopHeapEntry{
Name: string(name),
Count: count,
})
heap.Fix(th, len(th.a)-1)
return
}
if count <= th.a[0].Count {
return
}
th.a[0] = TopHeapEntry{
Name: string(name),
Count: count,
}
heap.Fix(th, 0)
}
func (th *topHeap) getSortedResult() []TopHeapEntry {
result := append([]TopHeapEntry{}, th.a...)
sort.Slice(result, func(i, j int) bool {
a, b := result[i], result[j]
if a.Count != b.Count {
return a.Count > b.Count
}
return a.Name < b.Name
})
return result
}
// heap.Interface implementation for topHeap.
func (th *topHeap) Len() int {
return len(th.a)
}
func (th *topHeap) Less(i, j int) bool {
a := th.a
return a[i].Count < a[j].Count
}
func (th *topHeap) Swap(i, j int) {
a := th.a
a[j], a[i] = a[i], a[j]
}
func (th *topHeap) Push(_ interface{}) {
panic(fmt.Errorf("BUG: Push shouldn't be called"))
}
func (th *topHeap) Pop() interface{} {
panic(fmt.Errorf("BUG: Pop shouldn't be called"))
}
// searchMetricNameWithCache appends metric name for the given metricID to dst
2019-05-22 21:16:55 +00:00
// and returns the result.
func (db *indexDB) searchMetricNameWithCache(dst []byte, metricID uint64, accountID, projectID uint32) ([]byte, bool) {
metricName := db.getMetricNameFromCache(dst, metricID)
if len(metricName) > len(dst) {
return metricName, true
}
is := db.getIndexSearch(accountID, projectID, noDeadline)
var ok bool
dst, ok = is.searchMetricName(dst, metricID)
2019-05-22 21:16:55 +00:00
db.putIndexSearch(is)
if ok {
// There is no need in verifying whether the given metricID is deleted,
// since the filtering must be performed before calling this func.
db.putMetricNameToCache(metricID, dst)
return dst, true
2019-05-22 21:16:55 +00:00
}
// Try searching in the external indexDB.
if db.doExtDB(func(extDB *indexDB) {
is := extDB.getIndexSearch(accountID, projectID, noDeadline)
dst, ok = is.searchMetricName(dst, metricID)
2019-05-22 21:16:55 +00:00
extDB.putIndexSearch(is)
if ok {
// There is no need in verifying whether the given metricID is deleted,
// since the filtering must be performed before calling this func.
extDB.putMetricNameToCache(metricID, dst)
}
}) && ok {
return dst, true
2019-05-22 21:16:55 +00:00
}
// Cannot find MetricName for the given metricID. This may be the case
// when indexDB contains incomplete set of metricID -> metricName entries
// after a snapshot or due to unflushed entries.
atomic.AddUint64(&db.missingMetricNamesForMetricID, 1)
// Mark the metricID as deleted, so it will be created again when new data point
// for the given time series will arrive.
db.deleteMetricIDs([]uint64{metricID})
return dst, false
2019-05-22 21:16:55 +00:00
}
// DeleteTSIDs marks as deleted all the TSIDs matching the given tfss.
//
// The caller must reset all the caches which may contain the deleted TSIDs.
//
// Returns the number of metrics deleted.
func (db *indexDB) DeleteTSIDs(qt *querytracer.Tracer, tfss []*TagFilters) (int, error) {
qt = qt.NewChild("deleting series for %s", tfss)
defer qt.Done()
2019-05-22 21:16:55 +00:00
if len(tfss) == 0 {
return 0, nil
}
// Obtain metricIDs to delete.
tr := TimeRange{
MinTimestamp: 0,
MaxTimestamp: (1 << 63) - 1,
}
is := db.getIndexSearch(tfss[0].accountID, tfss[0].projectID, noDeadline)
metricIDs, err := is.searchMetricIDs(qt, tfss, tr, 2e9)
2019-05-22 21:16:55 +00:00
db.putIndexSearch(is)
if err != nil {
return 0, err
}
db.deleteMetricIDs(metricIDs)
// Delete TSIDs in the extDB.
deletedCount := len(metricIDs)
db.doExtDB(func(extDB *indexDB) {
var n int
qtChild := qt.NewChild("deleting series from the previos indexdb")
n, err = extDB.DeleteTSIDs(qtChild, tfss)
qtChild.Donef("deleted %d series", n)
deletedCount += n
})
if err != nil {
return deletedCount, fmt.Errorf("cannot delete tsids in extDB: %w", err)
}
return deletedCount, nil
}
func (db *indexDB) deleteMetricIDs(metricIDs []uint64) {
2019-05-22 21:16:55 +00:00
if len(metricIDs) == 0 {
// Nothing to delete
return
2019-05-22 21:16:55 +00:00
}
// atomically add deleted metricIDs to an inmemory map.
dmis := &uint64set.Set{}
dmis.AddMulti(metricIDs)
db.s.updateDeletedMetricIDs(dmis)
2019-05-22 21:16:55 +00:00
// Reset TagFilters -> TSIDS cache, since it may contain deleted TSIDs.
invalidateTagFiltersCache()
2019-05-22 21:16:55 +00:00
// Reset MetricName -> TSID cache, since it may contain deleted TSIDs.
db.s.resetAndSaveTSIDCache()
// Store the metricIDs as deleted.
// Make this after updating the deletedMetricIDs and resetting caches
// in order to exclude the possibility of the inconsistent state when the deleted metricIDs
// remain available in the tsidCache after unclean shutdown.
// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1347
items := getIndexItems()
for _, metricID := range metricIDs {
items.B = append(items.B, nsPrefixDeletedMetricID)
items.B = encoding.MarshalUint64(items.B, metricID)
items.Next()
}
db.tb.AddItems(items.Items)
putIndexItems(items)
2019-05-22 21:16:55 +00:00
}
func (db *indexDB) loadDeletedMetricIDs() (*uint64set.Set, error) {
is := db.getIndexSearch(0, 0, noDeadline)
dmis, err := is.loadDeletedMetricIDs()
db.putIndexSearch(is)
if err != nil {
return nil, err
}
return dmis, nil
2019-05-22 21:16:55 +00:00
}
func (is *indexSearch) loadDeletedMetricIDs() (*uint64set.Set, error) {
dmis := &uint64set.Set{}
2019-05-22 21:16:55 +00:00
ts := &is.ts
kb := &is.kb
kb.B = append(kb.B[:0], nsPrefixDeletedMetricID)
2019-05-22 21:16:55 +00:00
ts.Seek(kb.B)
for ts.NextItem() {
item := ts.Item
if !bytes.HasPrefix(item, kb.B) {
break
}
item = item[len(kb.B):]
if len(item) != 8 {
return nil, fmt.Errorf("unexpected item len; got %d bytes; want %d bytes", len(item), 8)
}
metricID := encoding.UnmarshalUint64(item)
dmis.Add(metricID)
2019-05-22 21:16:55 +00:00
}
if err := ts.Error(); err != nil {
return nil, err
}
return dmis, nil
}
func (db *indexDB) searchMetricIDs(qt *querytracer.Tracer, tfss []*TagFilters, tr TimeRange, maxMetrics int, deadline uint64) ([]uint64, error) {
qt = qt.NewChild("search for matching metricIDs: filters=%s, timeRange=%s", tfss, &tr)
defer qt.Done()
2019-05-22 21:16:55 +00:00
if len(tfss) == 0 {
return nil, nil
}
if tr.MinTimestamp >= db.s.minTimestampForCompositeIndex {
tfss = convertToCompositeTagFilterss(tfss)
}
2019-05-22 21:16:55 +00:00
qtChild := qt.NewChild("search for metricIDs in the current indexdb")
2019-05-22 21:16:55 +00:00
tfKeyBuf := tagFiltersKeyBufPool.Get()
defer tagFiltersKeyBufPool.Put(tfKeyBuf)
tfKeyBuf.B = marshalTagFiltersKey(tfKeyBuf.B[:0], tfss, tr, true)
metricIDs, ok := db.getMetricIDsFromTagFiltersCache(qtChild, tfKeyBuf.B)
2019-05-22 21:16:55 +00:00
if ok {
// Fast path - metricIDs found in the cache
qtChild.Done()
return metricIDs, nil
2019-05-22 21:16:55 +00:00
}
// Slow path - search for metricIDs in the db and extDB.
accountID := tfss[0].accountID
projectID := tfss[0].projectID
is := db.getIndexSearch(accountID, projectID, deadline)
localMetricIDs, err := is.searchMetricIDs(qtChild, tfss, tr, maxMetrics)
2019-05-22 21:16:55 +00:00
db.putIndexSearch(is)
if err != nil {
return nil, fmt.Errorf("error when searching for metricIDs in the current indexdb: %s", err)
2019-05-22 21:16:55 +00:00
}
qtChild.Done()
2019-05-22 21:16:55 +00:00
var extMetricIDs []uint64
db.doExtDB(func(extDB *indexDB) {
qtChild := qt.NewChild("search for metricIDs in the previous indexdb")
defer qtChild.Done()
tfKeyExtBuf := tagFiltersKeyBufPool.Get()
defer tagFiltersKeyBufPool.Put(tfKeyExtBuf)
// Data in extDB cannot be changed, so use unversioned keys for tag cache.
tfKeyExtBuf.B = marshalTagFiltersKey(tfKeyExtBuf.B[:0], tfss, tr, false)
metricIDs, ok := extDB.getMetricIDsFromTagFiltersCache(qtChild, tfKeyExtBuf.B)
2019-05-22 21:16:55 +00:00
if ok {
extMetricIDs = metricIDs
2019-05-22 21:16:55 +00:00
return
}
is := extDB.getIndexSearch(accountID, projectID, deadline)
extMetricIDs, err = is.searchMetricIDs(qtChild, tfss, tr, maxMetrics)
2019-05-22 21:16:55 +00:00
extDB.putIndexSearch(is)
extDB.putMetricIDsToTagFiltersCache(qtChild, extMetricIDs, tfKeyExtBuf.B)
})
if err != nil {
return nil, fmt.Errorf("error when searching for metricIDs in the previous indexdb: %s", err)
2019-05-22 21:16:55 +00:00
}
// Merge localMetricIDs with extMetricIDs.
metricIDs = mergeSortedMetricIDs(localMetricIDs, extMetricIDs)
qt.Printf("merge %d metricIDs from the current indexdb with %d metricIDs from the previous indexdb; result: %d metricIDs",
len(localMetricIDs), len(extMetricIDs), len(metricIDs))
// Store metricIDs in the cache.
db.putMetricIDsToTagFiltersCache(qt, metricIDs, tfKeyBuf.B)
return metricIDs, nil
}
func mergeSortedMetricIDs(a, b []uint64) []uint64 {
if len(b) == 0 {
return a
}
i := 0
j := 0
result := make([]uint64, 0, len(a)+len(b))
for {
next := b[j]
start := i
for i < len(a) && a[i] <= next {
i++
}
result = append(result, a[start:i]...)
if len(result) > 0 {
last := result[len(result)-1]
for j < len(b) && b[j] == last {
j++
}
}
if i == len(a) {
return append(result, b[j:]...)
}
a, b = b, a
i, j = j, i
}
}
func (db *indexDB) getTSIDsFromMetricIDs(qt *querytracer.Tracer, accountID, projectID uint32, metricIDs []uint64, deadline uint64) ([]TSID, error) {
qt = qt.NewChild("obtain tsids from %d metricIDs", len(metricIDs))
defer qt.Done()
if len(metricIDs) == 0 {
return nil, nil
}
// Search for TSIDs in the current indexdb
tsids := make([]TSID, len(metricIDs))
var extMetricIDs []uint64
i := 0
err := func() error {
is := db.getIndexSearch(accountID, projectID, deadline)
defer db.putIndexSearch(is)
for loopsPaceLimiter, metricID := range metricIDs {
if loopsPaceLimiter&paceLimiterSlowIterationsMask == 0 {
if err := checkSearchDeadlineAndPace(is.deadline); err != nil {
return err
}
}
// Try obtaining TSIDs from MetricID->TSID cache. This is much faster
// than scanning the mergeset if it contains a lot of metricIDs.
tsid := &tsids[i]
err := is.db.getFromMetricIDCache(tsid, metricID)
if err == nil {
// Fast path - the tsid for metricID is found in cache.
i++
continue
}
if err != io.EOF {
return err
}
if !is.getTSIDByMetricID(tsid, metricID) {
// Postpone searching for the missing metricID in the extDB.
extMetricIDs = append(extMetricIDs, metricID)
continue
}
is.db.putToMetricIDCache(metricID, tsid)
i++
}
return nil
}()
if err != nil {
return nil, fmt.Errorf("error when searching for TISDs by metricIDs in the current indexdb: %w", err)
}
tsidsFound := i
qt.Printf("found %d tsids for %d metricIDs in the current indexdb", tsidsFound, len(metricIDs))
if len(extMetricIDs) > 0 {
// Search for extMetricIDs in the previous indexdb (aka extDB)
db.doExtDB(func(extDB *indexDB) {
is := extDB.getIndexSearch(accountID, projectID, deadline)
defer extDB.putIndexSearch(is)
for loopsPaceLimiter, metricID := range extMetricIDs {
if loopsPaceLimiter&paceLimiterSlowIterationsMask == 0 {
if err = checkSearchDeadlineAndPace(is.deadline); err != nil {
return
}
}
// There is no need in searching for TSIDs in MetricID->TSID cache, since
// this has been already done in the loop above (the MetricID->TSID cache is global).
tsid := &tsids[i]
if !is.getTSIDByMetricID(tsid, metricID) {
// Cannot find TSID for the given metricID.
// This may be the case on incomplete indexDB
// due to snapshot or due to unflushed entries.
// Just increment errors counter and skip it for now.
atomic.AddUint64(&is.db.missingTSIDsForMetricID, 1)
continue
}
is.db.putToMetricIDCache(metricID, tsid)
i++
}
})
if err != nil {
return nil, fmt.Errorf("error when searching for TSIDs by metricIDs in the previous indexdb: %w", err)
}
qt.Printf("found %d tsids for %d metricIDs in the previous indexdb", i-tsidsFound, len(extMetricIDs))
}
tsids = tsids[:i]
qt.Printf("load %d tsids for %d metricIDs from both current and previous indexdb", len(tsids), len(metricIDs))
2019-05-22 21:16:55 +00:00
// Sort the found tsids, since they must be passed to TSID search
// in the sorted order.
sort.Slice(tsids, func(i, j int) bool { return tsids[i].Less(&tsids[j]) })
qt.Printf("sort %d tsids", len(tsids))
return tsids, nil
2019-05-22 21:16:55 +00:00
}
var tagFiltersKeyBufPool bytesutil.ByteBufferPool
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
func (is *indexSearch) getTSIDByMetricNameNoExtDB(dst *TSID, metricName []byte, date uint64) bool {
dmis := is.db.s.getDeletedMetricIDs()
2019-05-22 21:16:55 +00:00
ts := &is.ts
kb := &is.kb
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
// Do not use marshalCommonPrefix() here, since mn already contains (AccountID, ProjectID)
kb.B = append(kb.B[:0], nsPrefixDateMetricNameToTSID)
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
kb.B = encoding.MarshalUint64(kb.B, date)
2019-05-22 21:16:55 +00:00
kb.B = append(kb.B, metricName...)
kb.B = append(kb.B, kvSeparatorChar)
ts.Seek(kb.B)
for ts.NextItem() {
if !bytes.HasPrefix(ts.Item, kb.B) {
// Nothing found.
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
return false
2019-05-22 21:16:55 +00:00
}
v := ts.Item[len(kb.B):]
tail, err := dst.Unmarshal(v)
if err != nil {
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
logger.Panicf("FATAL: cannot unmarshal TSID: %s", err)
2019-05-22 21:16:55 +00:00
}
if len(tail) > 0 {
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
logger.Panicf("FATAL: unexpected non-empty tail left after unmarshaling TSID: %X", tail)
2019-05-22 21:16:55 +00:00
}
if dmis.Len() > 0 {
2019-05-22 21:16:55 +00:00
// Verify whether the dst is marked as deleted.
if dmis.Has(dst.MetricID) {
2019-05-25 18:51:11 +00:00
// The dst is deleted. Continue searching.
2019-05-22 21:16:55 +00:00
continue
}
}
// Found valid dst.
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
return true
2019-05-22 21:16:55 +00:00
}
if err := ts.Error(); err != nil {
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
logger.Panicf("FATAL: error when searching TSID by metricName; searchPrefix %q: %s", kb.B, err)
2019-05-22 21:16:55 +00:00
}
// Nothing found
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
return false
2019-05-22 21:16:55 +00:00
}
func (is *indexSearch) searchMetricNameWithCache(dst []byte, metricID uint64) ([]byte, bool) {
2019-05-22 21:16:55 +00:00
metricName := is.db.getMetricNameFromCache(dst, metricID)
if len(metricName) > len(dst) {
return metricName, true
2019-05-22 21:16:55 +00:00
}
var ok bool
dst, ok = is.searchMetricName(dst, metricID)
if ok {
// There is no need in verifying whether the given metricID is deleted,
// since the filtering must be performed before calling this func.
is.db.putMetricNameToCache(metricID, dst)
return dst, true
}
return dst, false
}
2019-05-22 21:16:55 +00:00
func (is *indexSearch) searchMetricName(dst []byte, metricID uint64) ([]byte, bool) {
2019-05-22 21:16:55 +00:00
ts := &is.ts
kb := &is.kb
kb.B = is.marshalCommonPrefix(kb.B[:0], nsPrefixMetricIDToMetricName)
2019-05-22 21:16:55 +00:00
kb.B = encoding.MarshalUint64(kb.B, metricID)
if err := ts.FirstItemWithPrefix(kb.B); err != nil {
if err == io.EOF {
return dst, false
2019-05-22 21:16:55 +00:00
}
logger.Panicf("FATAL: error when searching metricName by metricID; searchPrefix %q: %w", kb.B, err)
2019-05-22 21:16:55 +00:00
}
v := ts.Item[len(kb.B):]
dst = append(dst, v...)
return dst, true
2019-05-22 21:16:55 +00:00
}
func (is *indexSearch) containsTimeRange(tr TimeRange) (bool, error) {
ts := &is.ts
kb := &is.kb
// Verify whether the tr.MinTimestamp is included into `ts` or is smaller than the minimum date stored in `ts`.
// Do not check whether tr.MaxTimestamp is included into `ts` or is bigger than the max date stored in `ts` for performance reasons.
// This means that containsTimeRange() can return true if `tr` is located below the min date stored in `ts`.
// This is OK, since this case isn't encountered too much in practice.
// The main practical case allows skipping searching in prev indexdb (`ts`) when `tr`
// is located above the max date stored there.
minDate := uint64(tr.MinTimestamp) / msecPerDay
kb.B = is.marshalCommonPrefix(kb.B[:0], nsPrefixDateToMetricID)
prefix := kb.B
kb.B = encoding.MarshalUint64(kb.B, minDate)
ts.Seek(kb.B)
if !ts.NextItem() {
if err := ts.Error(); err != nil {
return false, fmt.Errorf("error when searching for minDate=%d, prefix %q: %w", minDate, kb.B, err)
}
return false, nil
}
if !bytes.HasPrefix(ts.Item, prefix) {
// minDate exceeds max date from ts.
return false, nil
}
return true, nil
}
func (is *indexSearch) getTSIDByMetricID(dst *TSID, metricID uint64) bool {
2019-05-22 21:16:55 +00:00
// There is no need in checking for deleted metricIDs here, since they
// must be checked by the caller.
ts := &is.ts
kb := &is.kb
kb.B = is.marshalCommonPrefix(kb.B[:0], nsPrefixMetricIDToTSID)
2019-05-22 21:16:55 +00:00
kb.B = encoding.MarshalUint64(kb.B, metricID)
if err := ts.FirstItemWithPrefix(kb.B); err != nil {
if err == io.EOF {
return false
2019-05-22 21:16:55 +00:00
}
logger.Panicf("FATAL: error when searching TSID by metricID=%d; searchPrefix %q: %s", metricID, kb.B, err)
2019-05-22 21:16:55 +00:00
}
v := ts.Item[len(kb.B):]
tail, err := dst.Unmarshal(v)
if err != nil {
logger.Panicf("FATAL: cannot unmarshal the found TSID=%X for metricID=%d: %s", v, metricID, err)
2019-05-22 21:16:55 +00:00
}
if len(tail) > 0 {
logger.Panicf("FATAL: unexpected non-zero tail left after unmarshaling TSID for metricID=%d: %X", metricID, tail)
2019-05-22 21:16:55 +00:00
}
return true
2019-05-22 21:16:55 +00:00
}
// updateMetricIDsByMetricNameMatch matches metricName values for the given srcMetricIDs against tfs
2019-05-22 21:16:55 +00:00
// and adds matching metrics to metricIDs.
func (is *indexSearch) updateMetricIDsByMetricNameMatch(qt *querytracer.Tracer, metricIDs, srcMetricIDs *uint64set.Set, tfs []*tagFilter) error {
qt = qt.NewChild("filter out %d metric ids with filters=%s", srcMetricIDs.Len(), tfs)
defer qt.Done()
2019-05-22 21:16:55 +00:00
// sort srcMetricIDs in order to speed up Seek below.
sortedMetricIDs := srcMetricIDs.AppendTo(nil)
qt.Printf("sort %d metric ids", len(sortedMetricIDs))
2019-05-22 21:16:55 +00:00
kb := &is.kb
kb.B = is.marshalCommonPrefix(kb.B[:0], nsPrefixTagToMetricIDs)
tfs = removeCompositeTagFilters(tfs, kb.B)
2019-05-22 21:16:55 +00:00
metricName := kbPool.Get()
defer kbPool.Put(metricName)
mn := GetMetricName()
defer PutMetricName(mn)
for loopsPaceLimiter, metricID := range sortedMetricIDs {
if loopsPaceLimiter&paceLimiterSlowIterationsMask == 0 {
if err := checkSearchDeadlineAndPace(is.deadline); err != nil {
return err
}
}
var ok bool
metricName.B, ok = is.searchMetricNameWithCache(metricName.B[:0], metricID)
if !ok {
// It is likely the metricID->metricName entry didn't propagate to inverted index yet.
// Skip this metricID for now.
continue
2019-05-22 21:16:55 +00:00
}
if err := mn.Unmarshal(metricName.B); err != nil {
logger.Panicf("FATAL: cannot unmarshal metricName %q: %s", metricName.B, err)
2019-05-22 21:16:55 +00:00
}
// Match the mn against tfs.
ok, err := matchTagFilters(mn, tfs, &is.kb)
if err != nil {
return fmt.Errorf("cannot match MetricName %s against tagFilters: %w", mn, err)
2019-05-22 21:16:55 +00:00
}
if !ok {
continue
}
metricIDs.Add(metricID)
2019-05-22 21:16:55 +00:00
}
qt.Printf("apply filters %s; resulting metric ids: %d", tfs, metricIDs.Len())
2019-05-22 21:16:55 +00:00
return nil
}
func removeCompositeTagFilters(tfs []*tagFilter, prefix []byte) []*tagFilter {
if !hasCompositeTagFilters(tfs, prefix) {
return tfs
}
var tagKey []byte
var name []byte
tfsNew := make([]*tagFilter, 0, len(tfs)+1)
for _, tf := range tfs {
if !bytes.HasPrefix(tf.prefix, prefix) {
tfsNew = append(tfsNew, tf)
continue
}
suffix := tf.prefix[len(prefix):]
var err error
_, tagKey, err = unmarshalTagValue(tagKey[:0], suffix)
if err != nil {
logger.Panicf("BUG: cannot unmarshal tag key from suffix=%q: %s", suffix, err)
}
if len(tagKey) == 0 || tagKey[0] != compositeTagKeyPrefix {
tfsNew = append(tfsNew, tf)
continue
}
tagKey = tagKey[1:]
var nameLen uint64
tagKey, nameLen, err = encoding.UnmarshalVarUint64(tagKey)
if err != nil {
logger.Panicf("BUG: cannot unmarshal nameLen from tagKey %q: %s", tagKey, err)
}
if nameLen == 0 {
logger.Panicf("BUG: nameLen must be greater than 0")
}
if uint64(len(tagKey)) < nameLen {
logger.Panicf("BUG: expecting at %d bytes for name in tagKey=%q; got %d bytes", nameLen, tagKey, len(tagKey))
}
name = append(name[:0], tagKey[:nameLen]...)
tagKey = tagKey[nameLen:]
var tfNew tagFilter
if err := tfNew.Init(prefix, tagKey, tf.value, tf.isNegative, tf.isRegexp); err != nil {
logger.Panicf("BUG: cannot initialize {%s=%q} filter: %s", tagKey, tf.value, err)
}
tfsNew = append(tfsNew, &tfNew)
}
if len(name) > 0 {
var tfNew tagFilter
if err := tfNew.Init(prefix, nil, name, false, false); err != nil {
logger.Panicf("BUG: unexpected error when initializing {__name__=%q} filter: %s", name, err)
}
tfsNew = append(tfsNew, &tfNew)
}
return tfsNew
}
func hasCompositeTagFilters(tfs []*tagFilter, prefix []byte) bool {
var tagKey []byte
for _, tf := range tfs {
if !bytes.HasPrefix(tf.prefix, prefix) {
continue
}
suffix := tf.prefix[len(prefix):]
var err error
_, tagKey, err = unmarshalTagValue(tagKey[:0], suffix)
if err != nil {
logger.Panicf("BUG: cannot unmarshal tag key from suffix=%q: %s", suffix, err)
}
if len(tagKey) > 0 && tagKey[0] == compositeTagKeyPrefix {
return true
}
}
return false
}
2019-05-22 21:16:55 +00:00
func matchTagFilters(mn *MetricName, tfs []*tagFilter, kb *bytesutil.ByteBuffer) (bool, error) {
kb.B = marshalCommonPrefix(kb.B[:0], nsPrefixTagToMetricIDs, mn.AccountID, mn.ProjectID)
for i, tf := range tfs {
if bytes.Equal(tf.key, graphiteReverseTagKey) {
// Skip artificial tag filter for Graphite-like metric names with dots,
// since mn doesn't contain the corresponding tag.
continue
}
if len(tf.key) == 0 || string(tf.key) == "__graphite__" {
2019-05-22 21:16:55 +00:00
// Match against mn.MetricGroup.
b := marshalTagValue(kb.B, nil)
b = marshalTagValue(b, mn.MetricGroup)
kb.B = b[:len(kb.B)]
ok, err := tf.match(b)
2019-05-22 21:16:55 +00:00
if err != nil {
return false, fmt.Errorf("cannot match MetricGroup %q with tagFilter %s: %w", mn.MetricGroup, tf, err)
2019-05-22 21:16:55 +00:00
}
if !ok {
// Move failed tf to start.
// This should reduce the amount of useless work for the next mn.
if i > 0 {
tfs[0], tfs[i] = tfs[i], tfs[0]
}
2019-05-22 21:16:55 +00:00
return false, nil
}
continue
}
// Search for matching tag name.
tagMatched := false
tagSeen := false
for _, tag := range mn.Tags {
2019-05-22 21:16:55 +00:00
if string(tag.Key) != string(tf.key) {
continue
}
// Found the matching tag name. Match the value.
tagSeen = true
2019-05-22 21:16:55 +00:00
b := tag.Marshal(kb.B)
kb.B = b[:len(kb.B)]
ok, err := tf.match(b)
2019-05-22 21:16:55 +00:00
if err != nil {
return false, fmt.Errorf("cannot match tag %q with tagFilter %s: %w", tag, tf, err)
2019-05-22 21:16:55 +00:00
}
if !ok {
// Move failed tf to start.
// This should reduce the amount of useless work for the next mn.
if i > 0 {
tfs[0], tfs[i] = tfs[i], tfs[0]
}
2019-05-22 21:16:55 +00:00
return false, nil
}
tagMatched = true
break
}
if !tagSeen && (!tf.isNegative && tf.isEmptyMatch || tf.isNegative && !tf.isEmptyMatch) {
// tf contains positive empty-match filter for non-existing tag key, i.e.
// {non_existing_tag_key=~"foobar|"}
//
// OR
//
// tf contains negative filter for non-exsisting tag key
// and this filter doesn't match empty string, i.e. {non_existing_tag_key!="foobar"}
// Such filter matches anything.
//
// Note that the filter `{non_existing_tag_key!~"|foobar"}` shouldn't match anything,
// since it is expected that it matches non-empty `non_existing_tag_key`.
// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/546 and
// https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2255 for details.
continue
}
if tagMatched {
// tf matches mn. Go to the next tf.
continue
}
// Matching tag name wasn't found.
// Move failed tf to start.
// This should reduce the amount of useless work for the next mn.
if i > 0 {
tfs[0], tfs[i] = tfs[i], tfs[0]
2019-05-22 21:16:55 +00:00
}
return false, nil
2019-05-22 21:16:55 +00:00
}
return true, nil
}
func (is *indexSearch) searchMetricIDsWithFiltersOnDate(qt *querytracer.Tracer, tfss []*TagFilters, date uint64, maxMetrics int) (*uint64set.Set, error) {
if len(tfss) == 0 {
return nil, nil
}
tr := TimeRange{
MinTimestamp: int64(date) * msecPerDay,
MaxTimestamp: int64(date+1)*msecPerDay - 1,
}
if date == 0 {
// Search for metricIDs on the whole time range.
tr.MaxTimestamp = timestampFromTime(time.Now())
}
metricIDs, err := is.searchMetricIDsInternal(qt, tfss, tr, maxMetrics)
if err != nil {
return nil, err
}
return metricIDs, nil
}
func (is *indexSearch) searchMetricIDs(qt *querytracer.Tracer, tfss []*TagFilters, tr TimeRange, maxMetrics int) ([]uint64, error) {
ok, err := is.containsTimeRange(tr)
if err != nil {
return nil, err
}
if !ok {
// Fast path - the index doesn't contain data for the given tr.
return nil, nil
}
metricIDs, err := is.searchMetricIDsInternal(qt, tfss, tr, maxMetrics)
if err != nil {
return nil, err
2019-05-22 21:16:55 +00:00
}
if metricIDs.Len() == 0 {
2019-05-22 21:16:55 +00:00
// Nothing found
return nil, nil
}
sortedMetricIDs := metricIDs.AppendTo(nil)
qt.Printf("sort %d matching metric ids", len(sortedMetricIDs))
2019-05-22 21:16:55 +00:00
// Filter out deleted metricIDs.
dmis := is.db.s.getDeletedMetricIDs()
if dmis.Len() > 0 {
metricIDsFiltered := sortedMetricIDs[:0]
for _, metricID := range sortedMetricIDs {
if !dmis.Has(metricID) {
2019-05-22 21:16:55 +00:00
metricIDsFiltered = append(metricIDsFiltered, metricID)
}
}
qt.Printf("left %d metric ids after removing deleted metric ids", len(metricIDsFiltered))
sortedMetricIDs = metricIDsFiltered
2019-05-22 21:16:55 +00:00
}
return sortedMetricIDs, nil
2019-05-22 21:16:55 +00:00
}
func (is *indexSearch) searchMetricIDsInternal(qt *querytracer.Tracer, tfss []*TagFilters, tr TimeRange, maxMetrics int) (*uint64set.Set, error) {
qt = qt.NewChild("search for metric ids: filters=%s, timeRange=%s, maxMetrics=%d", tfss, &tr, maxMetrics)
defer qt.Done()
metricIDs := &uint64set.Set{}
for _, tfs := range tfss {
if len(tfs.tfs) == 0 {
// An empty filters must be equivalent to `{__name__!=""}`
tfs = NewTagFilters(tfs.accountID, tfs.projectID)
if err := tfs.Add(nil, nil, true, false); err != nil {
logger.Panicf(`BUG: cannot add {__name__!=""} filter: %s`, err)
}
}
qtChild := qt.NewChild("update metric ids: filters=%s, timeRange=%s", tfs, &tr)
prevMetricIDsLen := metricIDs.Len()
err := is.updateMetricIDsForTagFilters(qtChild, metricIDs, tfs, tr, maxMetrics+1)
qtChild.Donef("updated %d metric ids", metricIDs.Len()-prevMetricIDsLen)
if err != nil {
return nil, err
}
if metricIDs.Len() > maxMetrics {
return nil, fmt.Errorf("the number of matching timeseries exceeds %d; either narrow down the search "+
"or increase -search.max* command-line flag values at vmselect; see https://docs.victoriametrics.com/#resource-usage-limits", maxMetrics)
}
}
return metricIDs, nil
}
func (is *indexSearch) updateMetricIDsForTagFilters(qt *querytracer.Tracer, metricIDs *uint64set.Set, tfs *TagFilters, tr TimeRange, maxMetrics int) error {
err := is.tryUpdatingMetricIDsForDateRange(qt, metricIDs, tfs, tr, maxMetrics)
2020-03-13 20:42:22 +00:00
if err == nil {
// Fast path: found metricIDs by date range.
return nil
}
if !errors.Is(err, errFallbackToGlobalSearch) {
2020-03-13 20:42:22 +00:00
return err
}
// Slow path - fall back to search in the global inverted index.
qt.Printf("cannot find metric ids in per-day index; fall back to global index")
atomic.AddUint64(&is.db.globalSearchCalls, 1)
m, err := is.getMetricIDsForDateAndFilters(qt, 0, tfs, maxMetrics)
if err != nil {
if errors.Is(err, errFallbackToGlobalSearch) {
return fmt.Errorf("the number of matching timeseries exceeds %d; either narrow down the search "+
"or increase -search.max* command-line flag values at vmselect; see https://docs.victoriametrics.com/#resource-usage-limits", maxMetrics)
}
return err
2019-05-22 21:16:55 +00:00
}
metricIDs.UnionMayOwn(m)
2019-05-22 21:16:55 +00:00
return nil
}
func (is *indexSearch) getMetricIDsForTagFilter(qt *querytracer.Tracer, tf *tagFilter, maxMetrics int, maxLoopsCount int64) (*uint64set.Set, int64, error) {
2019-05-22 21:16:55 +00:00
if tf.isNegative {
logger.Panicf("BUG: isNegative must be false")
}
metricIDs := &uint64set.Set{}
2019-05-22 21:16:55 +00:00
if len(tf.orSuffixes) > 0 {
// Fast path for orSuffixes - seek for rows for each value from orSuffixes.
loopsCount, err := is.updateMetricIDsForOrSuffixes(tf, metricIDs, maxMetrics, maxLoopsCount)
qt.Printf("found %d metric ids for filter={%s} using exact search; spent %d loops", metricIDs.Len(), tf, loopsCount)
if err != nil {
return nil, loopsCount, fmt.Errorf("error when searching for metricIDs for tagFilter in fast path: %w; tagFilter=%s", err, tf)
2019-05-22 21:16:55 +00:00
}
return metricIDs, loopsCount, nil
2019-05-22 21:16:55 +00:00
}
// Slow path - scan for all the rows with the given prefix.
loopsCount, err := is.getMetricIDsForTagFilterSlow(tf, metricIDs.Add, maxLoopsCount)
qt.Printf("found %d metric ids for filter={%s} using prefix search; spent %d loops", metricIDs.Len(), tf, loopsCount)
if err != nil {
return nil, loopsCount, fmt.Errorf("error when searching for metricIDs for tagFilter in slow path: %w; tagFilter=%s", err, tf)
}
return metricIDs, loopsCount, nil
}
var errTooManyLoops = fmt.Errorf("too many loops is needed for applying this filter")
func (is *indexSearch) getMetricIDsForTagFilterSlow(tf *tagFilter, f func(metricID uint64), maxLoopsCount int64) (int64, error) {
if len(tf.orSuffixes) > 0 {
logger.Panicf("BUG: the getMetricIDsForTagFilterSlow must be called only for empty tf.orSuffixes; got %s", tf.orSuffixes)
}
// Scan all the rows with tf.prefix and call f on every tf match.
2019-05-22 21:16:55 +00:00
ts := &is.ts
kb := &is.kb
mp := &is.mp
var prevMatchingSuffix []byte
var prevMatch bool
var loopsCount int64
loopsPaceLimiter := 0
prefix := tf.prefix
ts.Seek(prefix)
for ts.NextItem() {
if loopsPaceLimiter&paceLimiterMediumIterationsMask == 0 {
if err := checkSearchDeadlineAndPace(is.deadline); err != nil {
return loopsCount, err
}
}
loopsPaceLimiter++
item := ts.Item
if !bytes.HasPrefix(item, prefix) {
return loopsCount, nil
2019-05-22 21:16:55 +00:00
}
tail := item[len(prefix):]
n := bytes.IndexByte(tail, tagSeparatorChar)
if n < 0 {
return loopsCount, fmt.Errorf("invalid tag->metricIDs line %q: cannot find tagSeparatorChar=%d", item, tagSeparatorChar)
2019-05-22 21:16:55 +00:00
}
suffix := tail[:n+1]
tail = tail[n+1:]
if err := mp.InitOnlyTail(item, tail); err != nil {
return loopsCount, err
2019-05-22 21:16:55 +00:00
}
mp.ParseMetricIDs()
loopsCount += int64(mp.MetricIDsLen())
if loopsCount > maxLoopsCount {
return loopsCount, errTooManyLoops
}
if prevMatch && string(suffix) == string(prevMatchingSuffix) {
// Fast path: the same tag value found.
// There is no need in checking it again with potentially
// slow tf.matchSuffix, which may call regexp.
for _, metricID := range mp.MetricIDs {
f(metricID)
}
continue
}
// Slow path: need tf.matchSuffix call.
ok, err := tf.matchSuffix(suffix)
// Assume that tf.matchSuffix call needs 10x more time than a single metric scan iteration.
loopsCount += 10 * int64(tf.matchCost)
2019-05-22 21:16:55 +00:00
if err != nil {
return loopsCount, fmt.Errorf("error when matching %s against suffix %q: %w", tf, suffix, err)
2019-05-22 21:16:55 +00:00
}
if !ok {
prevMatch = false
if mp.MetricIDsLen() < maxMetricIDsPerRow/2 {
// If the current row contains non-full metricIDs list,
// then it is likely the next row contains the next tag value.
// So skip seeking for the next tag value, since it will be slower than just ts.NextItem call.
continue
}
// Optimization: skip all the metricIDs for the given tag value
kb.B = append(kb.B[:0], item[:len(item)-len(tail)]...)
// The last char in kb.B must be tagSeparatorChar. Just increment it
// in order to jump to the next tag value.
if len(kb.B) == 0 || kb.B[len(kb.B)-1] != tagSeparatorChar || tagSeparatorChar >= 0xff {
return loopsCount, fmt.Errorf("data corruption: the last char in k=%X must be %X", kb.B, tagSeparatorChar)
}
kb.B[len(kb.B)-1]++
ts.Seek(kb.B)
// Assume that a seek cost is equivalent to 1000 ordinary loops.
loopsCount += 1000
2019-05-22 21:16:55 +00:00
continue
}
prevMatch = true
prevMatchingSuffix = append(prevMatchingSuffix[:0], suffix...)
for _, metricID := range mp.MetricIDs {
f(metricID)
}
2019-05-22 21:16:55 +00:00
}
if err := ts.Error(); err != nil {
return loopsCount, fmt.Errorf("error when searching for tag filter prefix %q: %w", prefix, err)
2019-05-22 21:16:55 +00:00
}
return loopsCount, nil
2019-05-22 21:16:55 +00:00
}
func (is *indexSearch) updateMetricIDsForOrSuffixes(tf *tagFilter, metricIDs *uint64set.Set, maxMetrics int, maxLoopsCount int64) (int64, error) {
2019-05-22 21:16:55 +00:00
if tf.isNegative {
logger.Panicf("BUG: isNegative must be false")
}
kb := kbPool.Get()
defer kbPool.Put(kb)
var loopsCount int64
2019-05-22 21:16:55 +00:00
for _, orSuffix := range tf.orSuffixes {
kb.B = append(kb.B[:0], tf.prefix...)
kb.B = append(kb.B, orSuffix...)
kb.B = append(kb.B, tagSeparatorChar)
lc, err := is.updateMetricIDsForOrSuffix(kb.B, metricIDs, maxMetrics, maxLoopsCount-loopsCount)
loopsCount += lc
if err != nil {
return loopsCount, err
2019-05-22 21:16:55 +00:00
}
if metricIDs.Len() >= maxMetrics {
return loopsCount, nil
2019-05-22 21:16:55 +00:00
}
}
return loopsCount, nil
2019-05-22 21:16:55 +00:00
}
func (is *indexSearch) updateMetricIDsForOrSuffix(prefix []byte, metricIDs *uint64set.Set, maxMetrics int, maxLoopsCount int64) (int64, error) {
2019-05-22 21:16:55 +00:00
ts := &is.ts
mp := &is.mp
var loopsCount int64
loopsPaceLimiter := 0
2019-05-22 21:16:55 +00:00
ts.Seek(prefix)
for metricIDs.Len() < maxMetrics && ts.NextItem() {
if loopsPaceLimiter&paceLimiterFastIterationsMask == 0 {
if err := checkSearchDeadlineAndPace(is.deadline); err != nil {
return loopsCount, err
}
}
loopsPaceLimiter++
item := ts.Item
if !bytes.HasPrefix(item, prefix) {
return loopsCount, nil
}
if err := mp.InitOnlyTail(item, item[len(prefix):]); err != nil {
return loopsCount, err
}
loopsCount += int64(mp.MetricIDsLen())
if loopsCount > maxLoopsCount {
return loopsCount, errTooManyLoops
}
mp.ParseMetricIDs()
metricIDs.AddMulti(mp.MetricIDs)
2019-05-22 21:16:55 +00:00
}
if err := ts.Error(); err != nil {
return loopsCount, fmt.Errorf("error when searching for tag filter prefix %q: %w", prefix, err)
2019-05-22 21:16:55 +00:00
}
return loopsCount, nil
2019-05-22 21:16:55 +00:00
}
2021-03-15 11:31:55 +00:00
var errFallbackToGlobalSearch = errors.New("fall back from per-day index search to global index search")
2019-05-22 21:16:55 +00:00
const maxDaysForPerDaySearch = 40
func (is *indexSearch) tryUpdatingMetricIDsForDateRange(qt *querytracer.Tracer, metricIDs *uint64set.Set, tfs *TagFilters, tr TimeRange, maxMetrics int) error {
atomic.AddUint64(&is.db.dateRangeSearchCalls, 1)
minDate := uint64(tr.MinTimestamp) / msecPerDay
maxDate := uint64(tr.MaxTimestamp-1) / msecPerDay
if minDate > maxDate || maxDate-minDate > maxDaysForPerDaySearch {
// Too much dates must be covered. Give up, since it may be slow.
2021-03-15 11:31:55 +00:00
return errFallbackToGlobalSearch
2020-03-13 20:42:22 +00:00
}
if minDate == maxDate {
// Fast path - query only a single date.
m, err := is.getMetricIDsForDateAndFilters(qt, minDate, tfs, maxMetrics)
2020-03-13 20:42:22 +00:00
if err != nil {
return err
}
metricIDs.UnionMayOwn(m)
2020-03-13 20:42:22 +00:00
atomic.AddUint64(&is.db.dateRangeSearchHits, 1)
return nil
}
2020-03-13 20:42:22 +00:00
// Slower path - search for metricIDs for each day in parallel.
qt = qt.NewChild("parallel search for metric ids in per-day index: filters=%s, dayRange=[%d..%d]", tfs, minDate, maxDate)
defer qt.Done()
wg := getWaitGroup()
var errGlobal error
2020-03-13 20:42:22 +00:00
var mu sync.Mutex // protects metricIDs + errGlobal vars from concurrent access below
for minDate <= maxDate {
qtChild := qt.NewChild("parallel thread for date=%s", dateToString(minDate))
wg.Add(1)
2020-03-13 20:42:22 +00:00
go func(date uint64) {
defer func() {
qtChild.Done()
wg.Done()
}()
isLocal := is.db.getIndexSearch(is.accountID, is.projectID, is.deadline)
m, err := isLocal.getMetricIDsForDateAndFilters(qtChild, date, tfs, maxMetrics)
is.db.putIndexSearch(isLocal)
mu.Lock()
2020-03-13 20:42:22 +00:00
defer mu.Unlock()
if errGlobal != nil {
return
}
if err != nil {
dateStr := time.Unix(int64(date*24*3600), 0)
2021-03-15 11:31:55 +00:00
errGlobal = fmt.Errorf("cannot search for metricIDs at %s: %w", dateStr, err)
2020-03-13 20:42:22 +00:00
return
}
2020-03-13 20:42:22 +00:00
if metricIDs.Len() < maxMetrics {
metricIDs.UnionMayOwn(m)
2020-03-13 20:42:22 +00:00
}
}(minDate)
minDate++
}
wg.Wait()
putWaitGroup(wg)
if errGlobal != nil {
2020-03-13 20:42:22 +00:00
return errGlobal
}
atomic.AddUint64(&is.db.dateRangeSearchHits, 1)
2020-03-13 20:42:22 +00:00
return nil
}
func (is *indexSearch) getMetricIDsForDateAndFilters(qt *querytracer.Tracer, date uint64, tfs *TagFilters, maxMetrics int) (*uint64set.Set, error) {
if qt.Enabled() {
qt = qt.NewChild("search for metric ids on a particular day: filters=%s, date=%s, maxMetrics=%d", tfs, dateToString(date), maxMetrics)
defer qt.Done()
}
// Sort tfs by loopsCount needed for performing each filter.
// This stats is usually collected from the previous queries.
// This way we limit the amount of work below by applying fast filters at first.
type tagFilterWithWeight struct {
tf *tagFilter
loopsCount int64
filterLoopsCount int64
}
tfws := make([]tagFilterWithWeight, len(tfs.tfs))
currentTime := fasttime.UnixTimestamp()
for i := range tfs.tfs {
tf := &tfs.tfs[i]
loopsCount, filterLoopsCount, timestamp := is.getLoopsCountAndTimestampForDateFilter(date, tf)
if currentTime > timestamp+3600 {
// Update stats once per hour for relatively fast tag filters.
// There is no need in spending CPU resources on updating stats for heavy tag filters.
if loopsCount <= 10e6 {
loopsCount = 0
}
if filterLoopsCount <= 10e6 {
filterLoopsCount = 0
}
2021-02-17 15:55:29 +00:00
}
tfws[i] = tagFilterWithWeight{
tf: tf,
loopsCount: loopsCount,
filterLoopsCount: filterLoopsCount,
}
}
sort.Slice(tfws, func(i, j int) bool {
a, b := &tfws[i], &tfws[j]
if a.loopsCount != b.loopsCount {
return a.loopsCount < b.loopsCount
}
return a.tf.Less(b.tf)
})
getFirstPositiveLoopsCount := func(tfws []tagFilterWithWeight) int64 {
for i := range tfws {
if n := tfws[i].loopsCount; n > 0 {
return n
}
}
return int64Max
}
storeLoopsCount := func(tfw *tagFilterWithWeight, loopsCount int64) {
if loopsCount != tfw.loopsCount {
tfw.loopsCount = loopsCount
is.storeLoopsCountForDateFilter(date, tfw.tf, tfw.loopsCount, tfw.filterLoopsCount)
}
}
// Populate metricIDs for the first non-negative filter with the smallest cost.
qtChild := qt.NewChild("search for the first non-negative filter with the smallest cost")
var metricIDs *uint64set.Set
tfwsRemaining := tfws[:0]
maxDateMetrics := intMax
if maxMetrics < intMax/50 {
maxDateMetrics = maxMetrics * 50
}
for i, tfw := range tfws {
tf := tfw.tf
if tf.isNegative || tf.isEmptyMatch {
tfwsRemaining = append(tfwsRemaining, tfw)
continue
}
maxLoopsCount := getFirstPositiveLoopsCount(tfws[i+1:])
m, loopsCount, err := is.getMetricIDsForDateTagFilter(qtChild, tf, date, tfs.commonPrefix, maxDateMetrics, maxLoopsCount)
if err != nil {
if errors.Is(err, errTooManyLoops) {
// The tf took too many loops compared to the next filter. Postpone applying this filter.
qtChild.Printf("the filter={%s} took more than %d loops; postpone it", tf, maxLoopsCount)
storeLoopsCount(&tfw, 2*loopsCount)
tfwsRemaining = append(tfwsRemaining, tfw)
continue
}
// Move failing filter to the end of filter list.
storeLoopsCount(&tfw, int64Max)
return nil, err
}
if m.Len() >= maxDateMetrics {
// Too many time series found by a single tag filter. Move the filter to the end of list.
qtChild.Printf("the filter={%s} matches at least %d series; postpone it", tf, maxDateMetrics)
storeLoopsCount(&tfw, int64Max-1)
tfwsRemaining = append(tfwsRemaining, tfw)
continue
}
storeLoopsCount(&tfw, loopsCount)
metricIDs = m
tfwsRemaining = append(tfwsRemaining, tfws[i+1:]...)
qtChild.Printf("the filter={%s} matches less than %d series (actually %d series); use it", tf, maxDateMetrics, metricIDs.Len())
break
}
qtChild.Done()
tfws = tfwsRemaining
if metricIDs == nil {
// All the filters in tfs are negative or match too many time series.
// Populate all the metricIDs for the given (date),
2020-03-13 20:42:22 +00:00
// so later they can be filtered out with negative filters.
qt.Printf("all the filters are negative or match more than %d time series; fall back to searching for all the metric ids", maxDateMetrics)
m, err := is.getMetricIDsForDate(date, maxDateMetrics)
2020-03-13 20:42:22 +00:00
if err != nil {
return nil, fmt.Errorf("cannot obtain all the metricIDs: %w", err)
}
if m.Len() >= maxDateMetrics {
// Too many time series found for the given (date). Fall back to global search.
2021-03-15 11:31:55 +00:00
return nil, errFallbackToGlobalSearch
}
2020-03-13 20:42:22 +00:00
metricIDs = m
qt.Printf("found %d metric ids", metricIDs.Len())
}
sort.Slice(tfws, func(i, j int) bool {
a, b := &tfws[i], &tfws[j]
if a.filterLoopsCount != b.filterLoopsCount {
return a.filterLoopsCount < b.filterLoopsCount
}
return a.tf.Less(b.tf)
})
getFirstPositiveFilterLoopsCount := func(tfws []tagFilterWithWeight) int64 {
for i := range tfws {
if n := tfws[i].filterLoopsCount; n > 0 {
return n
}
}
return int64Max
}
storeFilterLoopsCount := func(tfw *tagFilterWithWeight, filterLoopsCount int64) {
if filterLoopsCount != tfw.filterLoopsCount {
is.storeLoopsCountForDateFilter(date, tfw.tf, tfw.loopsCount, filterLoopsCount)
}
}
2020-03-13 20:42:22 +00:00
// Intersect metricIDs with the rest of filters.
//
// Do not run these tag filters in parallel, since this may result in CPU and RAM waste
2023-02-13 12:27:13 +00:00
// when the initial tag filters significantly reduce the number of found metricIDs,
// so the remaining filters could be performed via much faster metricName matching instead
// of slow selecting of matching metricIDs.
qtChild = qt.NewChild("intersect the remaining %d filters with the found %d metric ids", len(tfws), metricIDs.Len())
var tfsPostponed []*tagFilter
for i, tfw := range tfws {
tf := tfw.tf
metricIDsLen := metricIDs.Len()
if metricIDsLen == 0 {
// There is no need in applying the remaining filters to an empty set.
break
}
if tfw.filterLoopsCount > int64(metricIDsLen)*loopsCountPerMetricNameMatch {
// It should be faster performing metricName match on the remaining filters
// instead of scanning big number of entries in the inverted index for these filters.
for _, tfw := range tfws[i:] {
tfsPostponed = append(tfsPostponed, tfw.tf)
}
break
}
maxLoopsCount := getFirstPositiveFilterLoopsCount(tfws[i+1:])
if maxLoopsCount == int64Max {
maxLoopsCount = int64(metricIDsLen) * loopsCountPerMetricNameMatch
}
m, filterLoopsCount, err := is.getMetricIDsForDateTagFilter(qtChild, tf, date, tfs.commonPrefix, intMax, maxLoopsCount)
if err != nil {
if errors.Is(err, errTooManyLoops) {
// Postpone tf, since it took more loops than the next filter may need.
qtChild.Printf("postpone filter={%s}, since it took more than %d loops", tf, maxLoopsCount)
storeFilterLoopsCount(&tfw, 2*filterLoopsCount)
tfsPostponed = append(tfsPostponed, tf)
continue
}
// Move failing tf to the end of filter list
storeFilterLoopsCount(&tfw, int64Max)
return nil, err
}
storeFilterLoopsCount(&tfw, filterLoopsCount)
if tf.isNegative || tf.isEmptyMatch {
metricIDs.Subtract(m)
qtChild.Printf("subtract %d metric ids from the found %d metric ids for filter={%s}; resulting metric ids: %d", m.Len(), metricIDsLen, tf, metricIDs.Len())
} else {
metricIDs.Intersect(m)
qtChild.Printf("intersect %d metric ids with the found %d metric ids for filter={%s}; resulting metric ids: %d", m.Len(), metricIDsLen, tf, metricIDs.Len())
}
}
qtChild.Done()
if metricIDs.Len() == 0 {
// There is no need in applying tfsPostponed, since the result is empty.
qt.Printf("found zero metric ids")
return nil, nil
}
if len(tfsPostponed) > 0 {
// Apply the postponed filters via metricName match.
qt.Printf("apply postponed filters=%s to %d metrics ids", tfsPostponed, metricIDs.Len())
var m uint64set.Set
if err := is.updateMetricIDsByMetricNameMatch(qt, &m, metricIDs, tfsPostponed); err != nil {
return nil, err
}
return &m, nil
}
qt.Printf("found %d metric ids", metricIDs.Len())
2020-03-13 20:42:22 +00:00
return metricIDs, nil
}
const (
intMax = int((^uint(0)) >> 1)
int64Max = int64((1 << 63) - 1)
)
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
func (is *indexSearch) createPerDayIndexes(date uint64, tsid *TSID, mn *MetricName) {
ii := getIndexItems()
defer putIndexItems(ii)
ii.B = marshalCommonPrefix(ii.B, nsPrefixDateToMetricID, mn.AccountID, mn.ProjectID)
ii.B = encoding.MarshalUint64(ii.B, date)
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
ii.B = encoding.MarshalUint64(ii.B, tsid.MetricID)
ii.Next()
// Create per-day inverted index entries for TSID.
//
// Do not use marshalCommonPrefix() here, since mn already contains (AccountID, ProjectID)
ii.B = append(ii.B, nsPrefixDateMetricNameToTSID)
ii.B = encoding.MarshalUint64(ii.B, date)
ii.B = mn.Marshal(ii.B)
ii.B = append(ii.B, kvSeparatorChar)
ii.B = tsid.Marshal(ii.B)
ii.Next()
// Create per-day inverted index entries for metricID.
kb := kbPool.Get()
defer kbPool.Put(kb)
kb.B = marshalCommonPrefix(kb.B[:0], nsPrefixDateTagToMetricIDs, mn.AccountID, mn.ProjectID)
kb.B = encoding.MarshalUint64(kb.B, date)
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
ii.registerTagIndexes(kb.B, mn, tsid.MetricID)
is.db.tb.AddItems(ii.Items)
2019-05-22 21:16:55 +00:00
}
func (ii *indexItems) registerTagIndexes(prefix []byte, mn *MetricName, metricID uint64) {
// Add index entry for MetricGroup -> MetricID
ii.B = append(ii.B, prefix...)
ii.B = marshalTagValue(ii.B, nil)
ii.B = marshalTagValue(ii.B, mn.MetricGroup)
ii.B = encoding.MarshalUint64(ii.B, metricID)
ii.Next()
ii.addReverseMetricGroupIfNeeded(prefix, mn, metricID)
// Add index entries for tags: tag -> MetricID
for _, tag := range mn.Tags {
ii.B = append(ii.B, prefix...)
ii.B = tag.Marshal(ii.B)
ii.B = encoding.MarshalUint64(ii.B, metricID)
ii.Next()
}
// Add index entries for composite tags: MetricGroup+tag -> MetricID
compositeKey := kbPool.Get()
for _, tag := range mn.Tags {
compositeKey.B = marshalCompositeTagKey(compositeKey.B[:0], mn.MetricGroup, tag.Key)
ii.B = append(ii.B, prefix...)
ii.B = marshalTagValue(ii.B, compositeKey.B)
ii.B = marshalTagValue(ii.B, tag.Value)
ii.B = encoding.MarshalUint64(ii.B, metricID)
ii.Next()
}
kbPool.Put(compositeKey)
}
func (ii *indexItems) addReverseMetricGroupIfNeeded(prefix []byte, mn *MetricName, metricID uint64) {
if bytes.IndexByte(mn.MetricGroup, '.') < 0 {
// The reverse metric group is needed only for Graphite-like metrics with points.
return
}
// This is most likely a Graphite metric like 'foo.bar.baz'.
// Store reverse metric name 'zab.rab.oof' in order to speed up search for '*.bar.baz'
// when the Graphite wildcard has a suffix matching small number of time series.
ii.B = append(ii.B, prefix...)
ii.B = marshalTagValue(ii.B, graphiteReverseTagKey)
revBuf := kbPool.Get()
revBuf.B = reverseBytes(revBuf.B[:0], mn.MetricGroup)
ii.B = marshalTagValue(ii.B, revBuf.B)
kbPool.Put(revBuf)
ii.B = encoding.MarshalUint64(ii.B, metricID)
ii.Next()
}
func isArtificialTagKey(key []byte) bool {
if bytes.Equal(key, graphiteReverseTagKey) {
return true
}
if len(key) > 0 && key[0] == compositeTagKeyPrefix {
return true
}
return false
}
// The tag key for reverse metric name used for speeding up searching
// for Graphite wildcards with suffix matching small number of time series,
// i.e. '*.bar.baz'.
//
// It is expected that the given key isn't be used by users.
var graphiteReverseTagKey = []byte("\xff")
// The prefix for composite tag, which is used for speeding up searching
// for composite filters, which contain `{__name__="<metric_name>"}` filter.
//
// It is expected that the given prefix isn't used by users.
const compositeTagKeyPrefix = '\xfe'
func marshalCompositeTagKey(dst, name, key []byte) []byte {
dst = append(dst, compositeTagKeyPrefix)
dst = encoding.MarshalVarUint64(dst, uint64(len(name)))
dst = append(dst, name...)
dst = append(dst, key...)
return dst
}
func unmarshalCompositeTagKey(src []byte) ([]byte, []byte, error) {
if len(src) == 0 {
return nil, nil, fmt.Errorf("composite tag key cannot be empty")
}
if src[0] != compositeTagKeyPrefix {
return nil, nil, fmt.Errorf("missing composite tag key prefix in %q", src)
}
src = src[1:]
tail, n, err := encoding.UnmarshalVarUint64(src)
if err != nil {
return nil, nil, fmt.Errorf("cannot unmarshal metric name length from composite tag key: %w", err)
}
src = tail
if uint64(len(src)) < n {
return nil, nil, fmt.Errorf("missing metric name with length %d in composite tag key %q", n, src)
}
name := src[:n]
key := src[n:]
return name, key, nil
}
func reverseBytes(dst, src []byte) []byte {
for i := len(src) - 1; i >= 0; i-- {
dst = append(dst, src[i])
}
return dst
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
func (is *indexSearch) hasDateMetricIDNoExtDB(date, metricID uint64, accountID, projectID uint32) bool {
2019-05-22 21:16:55 +00:00
ts := &is.ts
kb := &is.kb
kb.B = marshalCommonPrefix(kb.B[:0], nsPrefixDateToMetricID, accountID, projectID)
2019-05-22 21:16:55 +00:00
kb.B = encoding.MarshalUint64(kb.B, date)
kb.B = encoding.MarshalUint64(kb.B, metricID)
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
err := ts.FirstItemWithPrefix(kb.B)
if err == nil {
if string(ts.Item) != string(kb.B) {
logger.Panicf("FATAL: unexpected entry for (date=%s, metricID=%d); got %q; want %q", dateToString(date), metricID, ts.Item, kb.B)
2019-05-22 21:16:55 +00:00
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
// Fast path - the (date, metricID) entry is found in the current indexdb.
return true
2019-05-22 21:16:55 +00:00
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
if err != io.EOF {
logger.Panicf("FATAL: unexpected error when searching for (date=%s, metricID=%d) entry: %s", dateToString(date), metricID, err)
2019-05-22 21:16:55 +00:00
}
lib/storage: switch from global to per-day index for `MetricName -> TSID` mapping Previously all the newly ingested time series were registered in global `MetricName -> TSID` index. This index was used during data ingestion for locating the TSID (internal series id) for the given canonical metric name (the canonical metric name consists of metric name plus all its labels sorted by label names). The `MetricName -> TSID` index is stored on disk in order to make sure that the data isn't lost on VictoriaMetrics restart or unclean shutdown. The lookup in this index is relatively slow, since VictoriaMetrics needs to read the corresponding data block from disk, unpack it, put the unpacked block into `indexdb/dataBlocks` cache, and then search for the given `MetricName -> TSID` entry there. So VictoriaMetrics uses in-memory cache for speeding up the lookup for active time series. This cache is named `storage/tsid`. If this cache capacity is enough for all the currently ingested active time series, then VictoriaMetrics works fast, since it doesn't need to read the data from disk. VictoriaMetrics starts reading data from `MetricName -> TSID` on-disk index in the following cases: - If `storage/tsid` cache capacity isn't enough for active time series. Then just increase available memory for VictoriaMetrics or reduce the number of active time series ingested into VictoriaMetrics. - If new time series is ingested into VictoriaMetrics. In this case it cannot find the needed entry in the `storage/tsid` cache, so it needs to consult on-disk `MetricName -> TSID` index, since it doesn't know that the index has no the corresponding entry too. This is a typical event under high churn rate, when old time series are constantly substituted with new time series. Reading the data from `MetricName -> TSID` index is slow, so inserts, which lead to reading this index, are counted as slow inserts, and they can be monitored via `vm_slow_row_inserts_total` metric exposed by VictoriaMetrics. Prior to this commit the `MetricName -> TSID` index was global, e.g. it contained entries sorted by `MetricName` for all the time series ever ingested into VictoriaMetrics during the configured -retentionPeriod. This index can become very large under high churn rate and long retention. VictoriaMetrics caches data from this index in `indexdb/dataBlocks` in-memory cache for speeding up index lookups. The `indexdb/dataBlocks` cache may occupy significant share of available memory for storing recently accessed blocks at `MetricName -> TSID` index when searching for newly ingested time series. This commit switches from global `MetricName -> TSID` index to per-day index. This allows significantly reducing the amounts of data, which needs to be cached in `indexdb/dataBlocks`, since now VictoriaMetrics consults only the index for the current day when new time series is ingested into it. The downside of this change is increased indexdb size on disk for workloads without high churn rate, e.g. with static time series, which do no change over time, since now VictoriaMetrics needs to store identical `MetricName -> TSID` entries for static time series for every day. This change removes an optimization for reducing CPU and disk IO spikes at indexdb rotation, since it didn't work correctly - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401 . At the same time the change fixes the issue, which could result in lost access to time series, which stop receving new samples during the first hour after indexdb rotation - see https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2698 The issue with the increased CPU and disk IO usage during indexdb rotation will be addressed in a separate commit according to https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1401#issuecomment-1553488685 This is a follow-up for 1f28b46ae9350795af41cbfc3ca0e8a5af084fce
2023-07-13 22:33:41 +00:00
return false
2019-05-22 21:16:55 +00:00
}
func (is *indexSearch) getMetricIDsForDateTagFilter(qt *querytracer.Tracer, tf *tagFilter, date uint64, commonPrefix []byte,
maxMetrics int, maxLoopsCount int64) (*uint64set.Set, int64, error) {
if qt.Enabled() {
qt = qt.NewChild("get metric ids for filter and date: filter={%s}, date=%s, maxMetrics=%d, maxLoopsCount=%d", tf, dateToString(date), maxMetrics, maxLoopsCount)
defer qt.Done()
}
if !bytes.HasPrefix(tf.prefix, commonPrefix) {
logger.Panicf("BUG: unexpected tf.prefix %q; must start with commonPrefix %q", tf.prefix, commonPrefix)
}
kb := kbPool.Get()
defer kbPool.Put(kb)
kb.B = is.marshalCommonPrefixForDate(kb.B[:0], date)
prefix := kb.B
kb.B = append(kb.B, tf.prefix[len(commonPrefix):]...)
tfNew := *tf
tfNew.isNegative = false // isNegative for the original tf is handled by the caller.
tfNew.prefix = kb.B
metricIDs, loopsCount, err := is.getMetricIDsForTagFilter(qt, &tfNew, maxMetrics, maxLoopsCount)
if err != nil {
return nil, loopsCount, err
}
if tf.isNegative || !tf.isEmptyMatch {
return metricIDs, loopsCount, nil
}
// The tag filter, which matches empty label such as {foo=~"bar|"}
// Convert it to negative filter, which matches {foo=~".+",foo!~"bar|"}.
// This fixes https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1601
// See also https://github.com/VictoriaMetrics/VictoriaMetrics/issues/395
maxLoopsCount -= loopsCount
var tfGross tagFilter
if err := tfGross.Init(prefix, tf.key, []byte(".+"), false, true); err != nil {
logger.Panicf(`BUG: cannot init tag filter: {%q=~".+"}: %s`, tf.key, err)
}
m, lc, err := is.getMetricIDsForTagFilter(qt, &tfGross, maxMetrics, maxLoopsCount)
loopsCount += lc
if err != nil {
return nil, loopsCount, err
}
mLen := m.Len()
m.Subtract(metricIDs)
qt.Printf("subtract %d metric ids for filter={%s} from %d metric ids for filter={%s}", metricIDs.Len(), &tfNew, mLen, &tfGross)
qt.Printf("found %d metric ids, spent %d loops", m.Len(), loopsCount)
return m, loopsCount, nil
}
func (is *indexSearch) getLoopsCountAndTimestampForDateFilter(date uint64, tf *tagFilter) (int64, int64, uint64) {
is.kb.B = appendDateTagFilterCacheKey(is.kb.B[:0], is.db.name, date, tf, is.accountID, is.projectID)
kb := kbPool.Get()
defer kbPool.Put(kb)
kb.B = is.db.loopsPerDateTagFilterCache.Get(kb.B[:0], is.kb.B)
if len(kb.B) != 3*8 {
return 0, 0, 0
}
loopsCount := encoding.UnmarshalInt64(kb.B)
filterLoopsCount := encoding.UnmarshalInt64(kb.B[8:])
timestamp := encoding.UnmarshalUint64(kb.B[16:])
return loopsCount, filterLoopsCount, timestamp
}
func (is *indexSearch) storeLoopsCountForDateFilter(date uint64, tf *tagFilter, loopsCount, filterLoopsCount int64) {
currentTimestamp := fasttime.UnixTimestamp()
is.kb.B = appendDateTagFilterCacheKey(is.kb.B[:0], is.db.name, date, tf, is.accountID, is.projectID)
kb := kbPool.Get()
kb.B = encoding.MarshalInt64(kb.B[:0], loopsCount)
kb.B = encoding.MarshalInt64(kb.B, filterLoopsCount)
kb.B = encoding.MarshalUint64(kb.B, currentTimestamp)
is.db.loopsPerDateTagFilterCache.Set(is.kb.B, kb.B)
kbPool.Put(kb)
}
func appendDateTagFilterCacheKey(dst []byte, indexDBName string, date uint64, tf *tagFilter, accountID, projectID uint32) []byte {
dst = append(dst, indexDBName...)
dst = encoding.MarshalUint64(dst, date)
dst = tf.Marshal(dst, accountID, projectID)
return dst
}
func (is *indexSearch) getMetricIDsForDate(date uint64, maxMetrics int) (*uint64set.Set, error) {
// Extract all the metricIDs from (date, __name__=value)->metricIDs entries.
kb := kbPool.Get()
defer kbPool.Put(kb)
kb.B = is.marshalCommonPrefixForDate(kb.B[:0], date)
kb.B = marshalTagValue(kb.B, nil)
2020-03-13 20:42:22 +00:00
var metricIDs uint64set.Set
if err := is.updateMetricIDsForPrefix(kb.B, &metricIDs, maxMetrics); err != nil {
return nil, err
}
return &metricIDs, nil
}
func (is *indexSearch) updateMetricIDsForPrefix(prefix []byte, metricIDs *uint64set.Set, maxMetrics int) error {
ts := &is.ts
mp := &is.mp
loopsPaceLimiter := 0
ts.Seek(prefix)
for ts.NextItem() {
if loopsPaceLimiter&paceLimiterFastIterationsMask == 0 {
if err := checkSearchDeadlineAndPace(is.deadline); err != nil {
return err
}
}
loopsPaceLimiter++
item := ts.Item
if !bytes.HasPrefix(item, prefix) {
return nil
2019-05-22 21:16:55 +00:00
}
tail := item[len(prefix):]
n := bytes.IndexByte(tail, tagSeparatorChar)
if n < 0 {
return fmt.Errorf("invalid tag->metricIDs line %q: cannot find tagSeparatorChar %d", item, tagSeparatorChar)
}
tail = tail[n+1:]
if err := mp.InitOnlyTail(item, tail); err != nil {
return err
}
mp.ParseMetricIDs()
metricIDs.AddMulti(mp.MetricIDs)
if metricIDs.Len() >= maxMetrics {
return nil
}
2019-05-22 21:16:55 +00:00
}
if err := ts.Error(); err != nil {
return fmt.Errorf("error when searching for all metricIDs by prefix %q: %w", prefix, err)
2019-05-22 21:16:55 +00:00
}
return nil
}
// The estimated number of index scan loops a single loop in updateMetricIDsByMetricNameMatch takes.
const loopsCountPerMetricNameMatch = 150
2019-05-22 21:16:55 +00:00
var kbPool bytesutil.ByteBufferPool
// Returns local unique MetricID.
func generateUniqueMetricID() uint64 {
// It is expected that metricIDs returned from this function must be dense.
// If they will be sparse, then this may hurt metric_ids intersection
// performance with uint64set.Set.
return atomic.AddUint64(&nextUniqueMetricID, 1)
2019-05-22 21:16:55 +00:00
}
// This number mustn't go backwards on restarts, otherwise metricID
// collisions are possible. So don't change time on the server
// between VictoriaMetrics restarts.
var nextUniqueMetricID = uint64(time.Now().UnixNano())
2019-05-22 21:16:55 +00:00
2019-05-22 21:23:23 +00:00
func marshalCommonPrefix(dst []byte, nsPrefix byte, accountID, projectID uint32) []byte {
2019-05-22 21:16:55 +00:00
dst = append(dst, nsPrefix)
2019-05-22 21:23:23 +00:00
dst = encoding.MarshalUint32(dst, accountID)
dst = encoding.MarshalUint32(dst, projectID)
2019-05-22 21:16:55 +00:00
return dst
}
func (is *indexSearch) marshalCommonPrefix(dst []byte, nsPrefix byte) []byte {
return marshalCommonPrefix(dst, nsPrefix, is.accountID, is.projectID)
}
func (is *indexSearch) marshalCommonPrefixForDate(dst []byte, date uint64) []byte {
if date == 0 {
// Global index
return is.marshalCommonPrefix(dst, nsPrefixTagToMetricIDs)
}
// Per-day index
dst = is.marshalCommonPrefix(dst, nsPrefixDateTagToMetricIDs)
return encoding.MarshalUint64(dst, date)
}
func unmarshalCommonPrefix(src []byte) ([]byte, byte, uint32, uint32, error) {
if len(src) < commonPrefixLen {
return nil, 0, 0, 0, fmt.Errorf("cannot unmarshal common prefix from %d bytes; need at least %d bytes; data=%X", len(src), commonPrefixLen, src)
}
prefix := src[0]
accountID := encoding.UnmarshalUint32(src[1:])
projectID := encoding.UnmarshalUint32(src[5:])
return src[commonPrefixLen:], prefix, accountID, projectID, nil
}
// 1 byte for prefix, 4 bytes for accountID, 4 bytes for projectID
const commonPrefixLen = 9
type tagToMetricIDsRowParser struct {
// NSPrefix contains the first byte parsed from the row after Init call.
// This is either nsPrefixTagToMetricIDs or nsPrefixDateTagToMetricIDs.
NSPrefix byte
// AccountID contains parsed value after Init call
AccountID uint32
// ProjectID contains parsed value after Init call
ProjectID uint32
// Date contains parsed date for nsPrefixDateTagToMetricIDs rows after Init call
Date uint64
// MetricIDs contains parsed MetricIDs after ParseMetricIDs call
MetricIDs []uint64
// metricIDsParsed is set to true after ParseMetricIDs call
metricIDsParsed bool
// Tag contains parsed tag after Init call
Tag Tag
// tail contains the remaining unparsed metricIDs
tail []byte
}
func (mp *tagToMetricIDsRowParser) Reset() {
mp.NSPrefix = 0
mp.AccountID = 0
mp.ProjectID = 0
mp.Date = 0
mp.MetricIDs = mp.MetricIDs[:0]
mp.metricIDsParsed = false
mp.Tag.Reset()
mp.tail = nil
}
// Init initializes mp from b, which should contain encoded tag->metricIDs row.
//
// b cannot be re-used until Reset call.
func (mp *tagToMetricIDsRowParser) Init(b []byte, nsPrefixExpected byte) error {
tail, nsPrefix, accountID, projectID, err := unmarshalCommonPrefix(b)
if err != nil {
return fmt.Errorf("invalid tag->metricIDs row %q: %w", b, err)
}
if nsPrefix != nsPrefixExpected {
return fmt.Errorf("invalid prefix for tag->metricIDs row %q; got %d; want %d", b, nsPrefix, nsPrefixExpected)
}
if nsPrefix == nsPrefixDateTagToMetricIDs {
// unmarshal date.
if len(tail) < 8 {
return fmt.Errorf("cannot unmarshal date from (date, tag)->metricIDs row %q from %d bytes; want at least 8 bytes", b, len(tail))
}
mp.Date = encoding.UnmarshalUint64(tail)
tail = tail[8:]
}
mp.NSPrefix = nsPrefix
mp.AccountID = accountID
mp.ProjectID = projectID
tail, err = mp.Tag.Unmarshal(tail)
if err != nil {
return fmt.Errorf("cannot unmarshal tag from tag->metricIDs row %q: %w", b, err)
}
return mp.InitOnlyTail(b, tail)
}
// MarshalPrefix marshals row prefix without tail to dst.
func (mp *tagToMetricIDsRowParser) MarshalPrefix(dst []byte) []byte {
dst = marshalCommonPrefix(dst, mp.NSPrefix, mp.AccountID, mp.ProjectID)
if mp.NSPrefix == nsPrefixDateTagToMetricIDs {
dst = encoding.MarshalUint64(dst, mp.Date)
}
dst = mp.Tag.Marshal(dst)
return dst
}
// InitOnlyTail initializes mp.tail from tail.
//
// b must contain tag->metricIDs row.
// b cannot be re-used until Reset call.
func (mp *tagToMetricIDsRowParser) InitOnlyTail(b, tail []byte) error {
if len(tail) == 0 {
return fmt.Errorf("missing metricID in the tag->metricIDs row %q", b)
}
if len(tail)%8 != 0 {
return fmt.Errorf("invalid tail length in the tag->metricIDs row; got %d bytes; must be multiple of 8 bytes", len(tail))
}
mp.tail = tail
mp.metricIDsParsed = false
return nil
}
// EqualPrefix returns true if prefixes for mp and x are equal.
//
// Prefix contains (tag, accountID, projectID)
func (mp *tagToMetricIDsRowParser) EqualPrefix(x *tagToMetricIDsRowParser) bool {
if !mp.Tag.Equal(&x.Tag) {
return false
}
return mp.ProjectID == x.ProjectID && mp.AccountID == x.AccountID && mp.Date == x.Date && mp.NSPrefix == x.NSPrefix
}
// MetricIDsLen returns the number of MetricIDs in the mp.tail
func (mp *tagToMetricIDsRowParser) MetricIDsLen() int {
return len(mp.tail) / 8
}
// ParseMetricIDs parses MetricIDs from mp.tail into mp.MetricIDs.
func (mp *tagToMetricIDsRowParser) ParseMetricIDs() {
if mp.metricIDsParsed {
return
}
tail := mp.tail
mp.MetricIDs = mp.MetricIDs[:0]
n := len(tail) / 8
if n <= cap(mp.MetricIDs) {
mp.MetricIDs = mp.MetricIDs[:n]
} else {
mp.MetricIDs = append(mp.MetricIDs[:cap(mp.MetricIDs)], make([]uint64, n-cap(mp.MetricIDs))...)
}
metricIDs := mp.MetricIDs
_ = metricIDs[n-1]
for i := 0; i < n; i++ {
if len(tail) < 8 {
logger.Panicf("BUG: tail cannot be smaller than 8 bytes; got %d bytes; tail=%X", len(tail), tail)
return
}
metricID := encoding.UnmarshalUint64(tail)
metricIDs[i] = metricID
tail = tail[8:]
}
mp.metricIDsParsed = true
}
// GetMatchingSeriesCount returns the number of series in mp, which match metricIDs from the given filter
// and do not match metricIDs from negativeFilter.
//
// if filter is empty, then all series in mp are taken into account.
func (mp *tagToMetricIDsRowParser) GetMatchingSeriesCount(filter, negativeFilter *uint64set.Set) int {
if filter == nil && negativeFilter.Len() == 0 {
return mp.MetricIDsLen()
}
mp.ParseMetricIDs()
n := 0
for _, metricID := range mp.MetricIDs {
if filter != nil && !filter.Has(metricID) {
continue
}
if !negativeFilter.Has(metricID) {
n++
}
}
return n
}
func mergeTagToMetricIDsRows(data []byte, items []mergeset.Item) ([]byte, []mergeset.Item) {
data, items = mergeTagToMetricIDsRowsInternal(data, items, nsPrefixTagToMetricIDs)
data, items = mergeTagToMetricIDsRowsInternal(data, items, nsPrefixDateTagToMetricIDs)
return data, items
}
func mergeTagToMetricIDsRowsInternal(data []byte, items []mergeset.Item, nsPrefix byte) ([]byte, []mergeset.Item) {
// Perform quick checks whether items contain rows starting from nsPrefix
// based on the fact that items are sorted.
if len(items) <= 2 {
// The first and the last row must remain unchanged.
return data, items
}
firstItem := items[0].Bytes(data)
if len(firstItem) > 0 && firstItem[0] > nsPrefix {
return data, items
}
lastItem := items[len(items)-1].Bytes(data)
if len(lastItem) > 0 && lastItem[0] < nsPrefix {
return data, items
}
// items contain at least one row starting from nsPrefix. Merge rows with common tag.
tmm := getTagToMetricIDsRowsMerger()
tmm.dataCopy = append(tmm.dataCopy[:0], data...)
tmm.itemsCopy = append(tmm.itemsCopy[:0], items...)
mp := &tmm.mp
mpPrev := &tmm.mpPrev
dstData := data[:0]
dstItems := items[:0]
for i, it := range items {
item := it.Bytes(data)
if len(item) == 0 || item[0] != nsPrefix || i == 0 || i == len(items)-1 {
// Write rows not starting with nsPrefix as-is.
// Additionally write the first and the last row as-is in order to preserve
2021-03-09 07:18:19 +00:00
// sort order for adjacent blocks.
dstData, dstItems = tmm.flushPendingMetricIDs(dstData, dstItems, mpPrev)
dstData = append(dstData, item...)
dstItems = append(dstItems, mergeset.Item{
Start: uint32(len(dstData) - len(item)),
End: uint32(len(dstData)),
})
continue
}
if err := mp.Init(item, nsPrefix); err != nil {
logger.Panicf("FATAL: cannot parse row starting with nsPrefix %d during merge: %s", nsPrefix, err)
}
if mp.MetricIDsLen() >= maxMetricIDsPerRow {
dstData, dstItems = tmm.flushPendingMetricIDs(dstData, dstItems, mpPrev)
dstData = append(dstData, item...)
dstItems = append(dstItems, mergeset.Item{
Start: uint32(len(dstData) - len(item)),
End: uint32(len(dstData)),
})
continue
}
if !mp.EqualPrefix(mpPrev) {
dstData, dstItems = tmm.flushPendingMetricIDs(dstData, dstItems, mpPrev)
}
mp.ParseMetricIDs()
tmm.pendingMetricIDs = append(tmm.pendingMetricIDs, mp.MetricIDs...)
mpPrev, mp = mp, mpPrev
if len(tmm.pendingMetricIDs) >= maxMetricIDsPerRow {
dstData, dstItems = tmm.flushPendingMetricIDs(dstData, dstItems, mpPrev)
}
}
if len(tmm.pendingMetricIDs) > 0 {
logger.Panicf("BUG: tmm.pendingMetricIDs must be empty at this point; got %d items: %d", len(tmm.pendingMetricIDs), tmm.pendingMetricIDs)
}
if !checkItemsSorted(dstData, dstItems) {
// Items could become unsorted if initial items contain duplicate metricIDs:
//
// item1: 1, 1, 5
// item2: 1, 4
//
// Items could become the following after the merge:
//
// item1: 1, 5
// item2: 1, 4
//
// i.e. item1 > item2
//
// Leave the original items unmerged, so they can be merged next time.
// This case should be quite rare - if multiple data points are simultaneously inserted
// into the same new time series from multiple concurrent goroutines.
atomic.AddUint64(&indexBlocksWithMetricIDsIncorrectOrder, 1)
dstData = append(dstData[:0], tmm.dataCopy...)
dstItems = append(dstItems[:0], tmm.itemsCopy...)
if !checkItemsSorted(dstData, dstItems) {
logger.Panicf("BUG: the original items weren't sorted; items=%q", dstItems)
}
}
putTagToMetricIDsRowsMerger(tmm)
atomic.AddUint64(&indexBlocksWithMetricIDsProcessed, 1)
return dstData, dstItems
}
var indexBlocksWithMetricIDsIncorrectOrder uint64
var indexBlocksWithMetricIDsProcessed uint64
func checkItemsSorted(data []byte, items []mergeset.Item) bool {
if len(items) == 0 {
return true
}
prevItem := items[0].String(data)
for _, it := range items[1:] {
currItem := it.String(data)
if prevItem > currItem {
return false
}
prevItem = currItem
}
return true
}
// maxMetricIDsPerRow limits the number of metricIDs in tag->metricIDs row.
//
// This reduces overhead on index and metaindex in lib/mergeset.
const maxMetricIDsPerRow = 64
type uint64Sorter []uint64
func (s uint64Sorter) Len() int { return len(s) }
func (s uint64Sorter) Less(i, j int) bool {
return s[i] < s[j]
}
func (s uint64Sorter) Swap(i, j int) {
s[i], s[j] = s[j], s[i]
}
type tagToMetricIDsRowsMerger struct {
pendingMetricIDs uint64Sorter
mp tagToMetricIDsRowParser
mpPrev tagToMetricIDsRowParser
itemsCopy []mergeset.Item
dataCopy []byte
}
func (tmm *tagToMetricIDsRowsMerger) Reset() {
tmm.pendingMetricIDs = tmm.pendingMetricIDs[:0]
tmm.mp.Reset()
tmm.mpPrev.Reset()
tmm.itemsCopy = tmm.itemsCopy[:0]
tmm.dataCopy = tmm.dataCopy[:0]
}
func (tmm *tagToMetricIDsRowsMerger) flushPendingMetricIDs(dstData []byte, dstItems []mergeset.Item, mp *tagToMetricIDsRowParser) ([]byte, []mergeset.Item) {
if len(tmm.pendingMetricIDs) == 0 {
// Nothing to flush
return dstData, dstItems
}
// Use sort.Sort instead of sort.Slice in order to reduce memory allocations.
sort.Sort(&tmm.pendingMetricIDs)
tmm.pendingMetricIDs = removeDuplicateMetricIDs(tmm.pendingMetricIDs)
// Marshal pendingMetricIDs
dstDataLen := len(dstData)
dstData = mp.MarshalPrefix(dstData)
for _, metricID := range tmm.pendingMetricIDs {
dstData = encoding.MarshalUint64(dstData, metricID)
}
dstItems = append(dstItems, mergeset.Item{
Start: uint32(dstDataLen),
End: uint32(len(dstData)),
})
tmm.pendingMetricIDs = tmm.pendingMetricIDs[:0]
return dstData, dstItems
}
func removeDuplicateMetricIDs(sortedMetricIDs []uint64) []uint64 {
if len(sortedMetricIDs) < 2 {
return sortedMetricIDs
}
prevMetricID := sortedMetricIDs[0]
hasDuplicates := false
for _, metricID := range sortedMetricIDs[1:] {
if prevMetricID == metricID {
hasDuplicates = true
break
}
prevMetricID = metricID
}
if !hasDuplicates {
return sortedMetricIDs
}
dstMetricIDs := sortedMetricIDs[:1]
prevMetricID = sortedMetricIDs[0]
for _, metricID := range sortedMetricIDs[1:] {
if prevMetricID == metricID {
continue
}
dstMetricIDs = append(dstMetricIDs, metricID)
prevMetricID = metricID
}
return dstMetricIDs
}
func getTagToMetricIDsRowsMerger() *tagToMetricIDsRowsMerger {
v := tmmPool.Get()
if v == nil {
return &tagToMetricIDsRowsMerger{}
}
return v.(*tagToMetricIDsRowsMerger)
}
func putTagToMetricIDsRowsMerger(tmm *tagToMetricIDsRowsMerger) {
tmm.Reset()
tmmPool.Put(tmm)
}
var tmmPool sync.Pool