VictoriaMetrics/lib/mergeset/part_search.go
Zakhar Bessarab 837d0d136d
lib/mergeset: add sparse indexdb cache (#7269)
Related issue:
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/7182

- add a separate index cache for searches which might read through large
amounts of random entries. Primary use-case for this is retention and
downsampling filters, when applying filters background merge needs to
fetch large amount of random entries which pollutes an index cache.
Using different caches allows to reduce effect on memory usage and cache
efficiency of the main cache while still having high cache hit rate. A
separate cache size is 5% of allowed memory.

- reduce size of indexdb/dataBlocks cache in order to free memory for
new sparse cache. Reduced size by 5% and moved this to a separate cache.

- add a separate metricName search which does not cache metric names -
this is needed in order to allow disabling metric name caching when
applying downsampling/retention filters. Applying filters during
background merge accesses random entries, this fills up cache and does
not provide an actual improvement due to random access nature.


Merge performance and memory usage stats before and after the change:

- before

![image](https://github.com/user-attachments/assets/485fffbb-c225-47ae-b5c5-bc8a7c57b36e)


- after

![image](https://github.com/user-attachments/assets/f4ba3440-7c1c-4ec1-bc54-4d2ab431eef5)

---------

Signed-off-by: Zakhar Bessarab <z.bessarab@victoriametrics.com>
2024-10-24 15:21:17 +02:00

371 lines
8.2 KiB
Go

package mergeset
import (
"fmt"
"io"
"sort"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/blockcache"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/bytesutil"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/encoding"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
)
type partSearch struct {
// Item contains the last item found after the call to NextItem.
//
// The Item content is valid until the next call to NextItem.
Item []byte
// p is a part to search.
p *part
// The remaining metaindex rows to scan, obtained from p.mrs.
mrs []metaindexRow
// The remaining block headers to scan in the current metaindexRow.
bhs []blockHeader
// err contains the last error.
err error
indexBuf []byte
compressedIndexBuf []byte
sb storageBlock
ib *inmemoryBlock
ibItemIdx int
sparse bool
}
func (ps *partSearch) reset() {
ps.Item = nil
ps.p = nil
ps.mrs = nil
ps.bhs = nil
ps.err = nil
ps.indexBuf = ps.indexBuf[:0]
ps.compressedIndexBuf = ps.compressedIndexBuf[:0]
ps.sb.Reset()
ps.ib = nil
ps.ibItemIdx = 0
}
// Init initializes ps for search in the p.
//
// Use Seek for search in p.
func (ps *partSearch) Init(p *part, sparse bool) {
ps.reset()
ps.p = p
ps.sparse = sparse
}
// Seek seeks for the first item greater or equal to k in ps.
func (ps *partSearch) Seek(k []byte) {
if err := ps.Error(); err != nil {
// Do nothing on unrecoverable error.
return
}
ps.err = nil
if string(k) > string(ps.p.ph.lastItem) {
// Not matching items in the part.
ps.err = io.EOF
return
}
if ps.tryFastSeek(k) {
return
}
ps.Item = nil
ps.mrs = ps.p.mrs
ps.bhs = nil
ps.indexBuf = ps.indexBuf[:0]
ps.compressedIndexBuf = ps.compressedIndexBuf[:0]
ps.sb.Reset()
ps.ib = nil
ps.ibItemIdx = 0
if string(k) <= string(ps.p.ph.firstItem) {
// The first item in the first block matches.
ps.err = ps.nextBlock()
return
}
// Locate the first metaindexRow to scan.
if len(ps.mrs) == 0 {
logger.Panicf("BUG: part without metaindex rows passed to partSearch")
}
n := sort.Search(len(ps.mrs), func(i int) bool {
return string(k) <= string(ps.mrs[i].firstItem)
})
if n > 0 {
// The given k may be located in the previous metaindexRow, so go to it.
n--
}
ps.mrs = ps.mrs[n:]
// Read block headers for the found metaindexRow.
if err := ps.nextBHS(); err != nil {
ps.err = err
return
}
// Locate the first block to scan.
n = sort.Search(len(ps.bhs), func(i int) bool {
return string(k) <= string(ps.bhs[i].firstItem)
})
if n > 0 {
// The given k may be located in the previous block, so go to it.
n--
}
ps.bhs = ps.bhs[n:]
// Read the block.
if err := ps.nextBlock(); err != nil {
ps.err = err
return
}
// Locate the first item to scan in the block.
items := ps.ib.items
data := ps.ib.data
cpLen := commonPrefixLen(ps.ib.commonPrefix, k)
ps.ibItemIdx = binarySearchKey(data, items, k, cpLen)
if ps.ibItemIdx < len(items) {
// The item has been found.
return
}
// Nothing found in the current block. Proceed to the next block.
// The item to search must be the first in the next block.
if err := ps.nextBlock(); err != nil {
ps.err = err
return
}
}
func (ps *partSearch) tryFastSeek(k []byte) bool {
if ps.ib == nil {
return false
}
items := ps.ib.items
idx := ps.ibItemIdx
if idx >= len(items) {
// The ib is exhausted.
return false
}
cpLen := commonPrefixLen(ps.ib.commonPrefix, k)
suffix := k[cpLen:]
it := items[len(items)-1]
it.Start += uint32(cpLen)
data := ps.ib.data
if string(suffix) > it.String(data) {
// The item is located in next blocks.
return false
}
// The item is located either in the current block or in previous blocks.
if idx > 0 {
idx--
}
it = items[idx]
it.Start += uint32(cpLen)
if string(suffix) < it.String(data) {
items = items[:idx]
if len(items) == 0 {
return false
}
it = items[0]
it.Start += uint32(cpLen)
if string(suffix) < it.String(data) {
// The item is located in previous blocks.
return false
}
idx = 0
}
// The item is located in the current block
ps.ibItemIdx = idx + binarySearchKey(data, items[idx:], k, cpLen)
return true
}
// NextItem advances to the next Item.
//
// Returns true on success.
func (ps *partSearch) NextItem() bool {
if ps.err != nil {
return false
}
items := ps.ib.items
if ps.ibItemIdx < len(items) {
// Fast path - the current block contains more items.
// Proceed to the next item.
ps.Item = items[ps.ibItemIdx].Bytes(ps.ib.data)
ps.ibItemIdx++
return true
}
// The current block is over. Proceed to the next block.
if err := ps.nextBlock(); err != nil {
if err != io.EOF {
err = fmt.Errorf("error in %q: %w", ps.p.path, err)
}
ps.err = err
return false
}
// Invariant: len(ps.ib.items) > 0 after nextBlock.
ps.Item = ps.ib.items[0].Bytes(ps.ib.data)
ps.ibItemIdx++
return true
}
// Error returns the last error occurred in the ps.
func (ps *partSearch) Error() error {
if ps.err == io.EOF {
return nil
}
return ps.err
}
func (ps *partSearch) nextBlock() error {
if len(ps.bhs) == 0 {
// The current metaindexRow is over. Proceed to the next metaindexRow.
if err := ps.nextBHS(); err != nil {
return err
}
}
bh := &ps.bhs[0]
ps.bhs = ps.bhs[1:]
ib, err := ps.getInmemoryBlock(bh)
if err != nil {
return err
}
ps.ib = ib
ps.ibItemIdx = 0
return nil
}
func (ps *partSearch) nextBHS() error {
if len(ps.mrs) == 0 {
return io.EOF
}
mr := &ps.mrs[0]
ps.mrs = ps.mrs[1:]
idxbKey := blockcache.Key{
Part: ps.p,
Offset: mr.indexBlockOffset,
}
b := idxbCache.GetBlock(idxbKey)
if b == nil {
idxb, err := ps.readIndexBlock(mr)
if err != nil {
return fmt.Errorf("cannot read index block: %w", err)
}
b = idxb
idxbCache.PutBlock(idxbKey, b)
}
idxb := b.(*indexBlock)
ps.bhs = idxb.bhs
return nil
}
func (ps *partSearch) readIndexBlock(mr *metaindexRow) (*indexBlock, error) {
ps.compressedIndexBuf = bytesutil.ResizeNoCopyMayOverallocate(ps.compressedIndexBuf, int(mr.indexBlockSize))
ps.p.indexFile.MustReadAt(ps.compressedIndexBuf, int64(mr.indexBlockOffset))
var err error
ps.indexBuf, err = encoding.DecompressZSTD(ps.indexBuf[:0], ps.compressedIndexBuf)
if err != nil {
return nil, fmt.Errorf("cannot decompress index block: %w", err)
}
idxb := &indexBlock{
buf: append([]byte{}, ps.indexBuf...),
}
idxb.bhs, err = unmarshalBlockHeadersNoCopy(idxb.bhs[:0], idxb.buf, int(mr.blockHeadersCount))
if err != nil {
return nil, fmt.Errorf("cannot unmarshal block headers from index block (offset=%d, size=%d): %w", mr.indexBlockOffset, mr.indexBlockSize, err)
}
return idxb, nil
}
func (ps *partSearch) getInmemoryBlock(bh *blockHeader) (*inmemoryBlock, error) {
cache := ibCache
if ps.sparse {
cache = ibSparseCache
}
ibKey := blockcache.Key{
Part: ps.p,
Offset: bh.itemsBlockOffset,
}
b := cache.GetBlock(ibKey)
if b == nil {
ib, err := ps.readInmemoryBlock(bh)
if err != nil {
return nil, err
}
b = ib
cache.PutBlock(ibKey, b)
}
ib := b.(*inmemoryBlock)
return ib, nil
}
func (ps *partSearch) readInmemoryBlock(bh *blockHeader) (*inmemoryBlock, error) {
ps.sb.Reset()
ps.sb.itemsData = bytesutil.ResizeNoCopyMayOverallocate(ps.sb.itemsData, int(bh.itemsBlockSize))
ps.p.itemsFile.MustReadAt(ps.sb.itemsData, int64(bh.itemsBlockOffset))
ps.sb.lensData = bytesutil.ResizeNoCopyMayOverallocate(ps.sb.lensData, int(bh.lensBlockSize))
ps.p.lensFile.MustReadAt(ps.sb.lensData, int64(bh.lensBlockOffset))
ib := &inmemoryBlock{}
if err := ib.UnmarshalData(&ps.sb, bh.firstItem, bh.commonPrefix, bh.itemsCount, bh.marshalType); err != nil {
return nil, fmt.Errorf("cannot unmarshal storage block with %d items: %w", bh.itemsCount, err)
}
return ib, nil
}
func binarySearchKey(data []byte, items []Item, k []byte, cpLen int) int {
if len(items) == 0 {
return 0
}
suffix := k[cpLen:]
it := items[0]
it.Start += uint32(cpLen)
if string(suffix) <= it.String(data) {
// Fast path - the item is the first.
return 0
}
items = items[1:]
offset := uint(1)
// This has been copy-pasted from https://golang.org/src/sort/search.go
n := uint(len(items))
i, j := uint(0), n
for i < j {
h := uint(i+j) >> 1
it := items[h]
it.Start += uint32(cpLen)
if h >= 0 && h < uint(len(items)) && string(suffix) > it.String(data) {
i = h + 1
} else {
j = h
}
}
return int(i + offset)
}