### Describe Your Changes
Previously, "selector @ another_selector" assumed that
"another_selector" metric is supposed to exist since "start" used in the
query.
If the query was evaluated in the following case (timestamps):
- start - 2, end - 10
- "another_selector" 5,6,7,8,9,10
- "selector" The resulting "at" timestamp would be taken from NaN (as
`int64(NaN * 1000)`), causing a panic or invalid behavior later.
Note that type cast of `NaN` to int64 is also platform-dependent, so
value of `int64(math.NaN() * 1000)` can produce `0` or max int64 on
different platforms and versions of Go.
This commit changes this and checks for the first non-NaN value. This
makes it easier to use for users as series are not always aligned and
returning an error in this case would disallow using this for some time
ranges.
See: https://github.com/VictoriaMetrics/VictoriaMetrics/issues/8444
### Checklist
The following checks are **mandatory**:
- [x] My change adheres [VictoriaMetrics contributing
guidelines](https://docs.victoriametrics.com/contributing/).
---------
Signed-off-by: Zakhar Bessarab <z.bessarab@victoriametrics.com>
Signed-off-by: hagen1778 <roman@victoriametrics.com>
Co-authored-by: hagen1778 <roman@victoriametrics.com>
(cherry picked from commit 7dfaef9088)
### Describe Your Changes
#8342
fix negative rate result when the lookbehind window is longer than
`-search.maxLookback` or `-search.maxStalenessInterval` and data
contains gap.
This issue was introduced since
[v1.110.0](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/8072).
### Checklist
The following checks are **mandatory**:
- [x] My change adheres [VictoriaMetrics contributing
guidelines](https://docs.victoriametrics.com/contributing/).
See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/8072
### Describe Your Changes
Please provide a brief description of the changes you made. Be as
specific as possible to help others understand the purpose and impact of
your modifications.
### Checklist
The following checks are **mandatory**:
- [ ] My change adheres [VictoriaMetrics contributing
guidelines](https://docs.victoriametrics.com/contributing/).
---------
Signed-off-by: hagen1778 <roman@victoriametrics.com>
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
`doInternal` has adaptive staleness detection mechanism. It is
calculated using timestamp distance between samples in selected list of
samples. It is dynamic because VM can store signals from many sources
with different samples resolution. And while it works for most of cases,
there are edge cases for rollup functions that are comparing values
between windows: increase, increase_pure, delta.
The edge case 1.
There was a gap between series because of the missed scrape or two. In
this case staleness will trigger and increase-like functions will assume
the value they need to compare with is 0. In result, this could produce
spikes for a flappy data - see
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/894 This
problem was solved by introducing a `realPrevValue` field -
1f19c167a4.
It stores the closest real sample value on selected interval and is used
for comparison when samples start after the gap.
The edge case 2.
`realPrevValue` doesn't respect staleness interval. In result, even if
gap between samples is huge (hours), the increase-like functions will
not consider it as a new series that started from 0. See
https://github.com/VictoriaMetrics/VictoriaMetrics/pull/8002.
Covering both edge cases is tricky, because `realPrevValue` has to
respect and not respect the staleness interval in the same time. In
other words, it should be able to ignore periodic missing gaps, but
reset if the gap is too big. While "too big gap" can't be figured out
empirically, I suggest using `-search.maxStalenessInterval` for this
purpose. If `-search.maxStalenessInterval` is set to 0 (default), then
`realPrevValue` ignores staleness interval. If
`-search.maxStalenessInterval` is > 0, then `realPrevValue` respects it
as a staleness interval.
### Checklist
The following checks are **mandatory**:
- [ ] My change adheres [VictoriaMetrics contributing
guidelines](https://docs.victoriametrics.com/contributing/).
---------
Signed-off-by: hagen1778 <roman@victoriametrics.com>
(cherry picked from commit 7d2a6764e7)
New version has additional checks and reduced resource consumption, so
it doesn't timeout for our internal repos.
To make linter happy, I addressed "redefinition of the built-in
function" lint error.
----
Signed-off-by: hagen1778 <roman@victoriametrics.com>
When vmselect process a rollup function it fetches all the raw samples
on requested `start-end` interval of the query. It then loops through
the raw samples, picks the range of the samples based on provided `step`
interval and invokes a rollup function for each of the picked ranges of
samples.
During this processing, vmselect always populates the `realPrevValue`
field with the closest previous raw sample value before the picked range
of samples. This `realPrevValue` is used by rollup functions like
increase_pure or delta to decide whether the counter change happened or
not. For example, we get the counter value == 1. If we've seen this
counter before and its value was also 1 - then no change happened. If we
didn't see it before, then this counter should have started with value=0
and we need to account for `1-0=1` change. All this is required to deal
with situations when scrapes are missing or `step` is too small.
However, vmselect doesn't check how "old" is the `realPrevValue`. In
other words, it doesn't respect the staleness interval when picking it.
In result, depending on the `start` and `end` params, vmselect can use
`realPrevValue` which is a couple of hours old and is unlikely to be a
temporary scrape fail. In result, some increases can be incorrectly
ingnored by vmselect.
This change makes sure that vmselect doesn't populate `realPrevValue`
with samples that are older than staleness interval.
### Describe Your Changes
Please provide a brief description of the changes you made. Be as
specific as possible to help others understand the purpose and impact of
your modifications.
### Checklist
The following checks are **mandatory**:
- [ x ] My change adheres [VictoriaMetrics contributing
guidelines](https://docs.victoriametrics.com/contributing/).
-------------------
To reproduce, create a dataset with one metric `foo` which has samples
with value=1 on interval of couple of hours and resolution 15s, and a
gap for an hour in the middle:
<img width="769" alt="image"
src="https://github.com/user-attachments/assets/a39b2740-b741-45f8-ad18-093b7c57c3b3"
/>
Then run `increase(foo[1m])` expression on this time range (disable
cache):
<img width="1472" alt="image"
src="https://github.com/user-attachments/assets/463cece1-f359-4c75-a96c-60092a31cab2"
/>
In result, there will be one increase on the beginning of the series.
And no increase after the gap. Then change the time range so it starts
in the middle of the gap:
<img width="1505" alt="image"
src="https://github.com/user-attachments/assets/f4a460c3-9fd1-4ec7-ab47-15e716ec1019"
/>
Now, there is an increase>0 because the `realPrevValue` wasn't
populated. This is wrong, because it hides the increase of the series.
With the fix, the original increase query on full time range should show
2 increases:
<img width="1492" alt="image"
src="https://github.com/user-attachments/assets/aa9d8a6b-7b22-41f6-9eb9-83b3113a6982"
/>
Signed-off-by: hagen1778 <roman@victoriametrics.com>
Since
44b071296d
`evalNumber` function no longer updating MetricName tenancy information.
This leads to mismatch in metric names between the query result and
evaluated number for all tenants other than 0:0.
For example, query `count(up) or 0` will return different results for
tenants 0:0 and 1:1 (assuming up is present for both tenants):
- tenant 0:0 - will only contain result of `count(up)`
- tenant 1:1 - will return both `count(up)` and `0` since metric names
will not be matched
This restores setting of tenancy information for metric name for
single-tenant queries.
Related issue:
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/7987
---
Signed-off-by: Zakhar Bessarab <z.bessarab@victoriametrics.com>
### Describe Your Changes
Binary operations like `exprFirst op exprSecond` in VictoriaMetrics are
performed in the following way:
1. Execute exprFirst.
2. Extract **common label filters** from the result of step 1.
3. Apply these common label filters to `exprSecond` and execute it, in
order to retrieve less time series from vmstorage nodes.
In step 2, only labels with less than `100` (hard-coded) value could be
used as **common label filter** (e.g. `{common_lb=~"v1|v2|...|v100"}`.
In our scenarios, a label, take `instance` label as an example, could
has thousands of candidate values. Regarding bring more pressure to
vmstorage node, it's still beneficial if labels with more than 100
values could be used as filter in `exprSecond`, with enough vmstorage
resources. After adjusting the value from `100` to `10000`, our query
round-trip time drops significantly from 5s to 2s.
This pull request change the hard-coded value into a configurable flag.
Parse cache is a pretty simple implementation of cache. It's just a
standard map with mutex.
Map with mutex overall has poor performance, plus when the cache
overflow occurs, the whole cache locks until 1k elements have been
deleted (now it's 10% of 10000 max elements in the cache). To avoid this
bottleneck and improve performance of cache on systems with many CPU
cores but keep it rather simple, we can implement cache with per bucket
locks like it's done in fastcache. The logic and API remain the same. So
now each bucket will have a map with approximately 78 elements (with 128
buckets), and overflow will occur now for each bucket, and only 7
elements need to be deleted.
Because exec_test.go has about 10k lines of code, it's better to move
the cache into a separate file to add tests and benchmarks for it,
because now it does not have them.
```
goos: windows
goarch: amd64
pkg: github.com/VictoriaMetrics/VictoriaMetrics/app/vmselect/promql
cpu: 11th Gen Intel(R) Core(TM) i9-11900K @ 3.50GHz
Current cache implementation performance on 8 cores:
BenchmarkCachePutNoOverFlow-8 1932 618372 ns/op 253 B/op 0 allocs/op
BenchmarkCacheGetNoOverflow-8 6547 211527 ns/op 0 B/op 0 allocs/op
BenchmarkCachePutGetNoOverflow-8 1873 621718 ns/op 261 B/op 0 allocs/op
BenchmarkCachePutOverflow-8 2262 464328 ns/op 32 B/op 0 allocs/op
BenchmarkCachePutGetOverflow-8 1764 655866 ns/op 38 B/op 0 allocs/op
New cache implementation performance on 8 cores:
BenchmarkCachePutNoOverFlow-8 10408 111412 ns/op 0 B/op 0 allocs/op
BenchmarkCacheGetNoOverflow-8 22407 52809 ns/op 0 B/op 0 allocs/op
BenchmarkCachePutGetNoOverflow-8 6583 168088 ns/op 0 B/op 0 allocs/op
BenchmarkCachePutOverflow-8 9822 117212 ns/op 2 B/op 0 allocs/op
BenchmarkCachePutGetOverflow-8 6481 175952 ns/op 3 B/op 0 allocs/op
Current cache implementation performance on 16 cores:
BenchmarkCachePutNoOverFlow-16 2331 475307 ns/op 218 B/op 0 allocs/op
BenchmarkCacheGetNoOverflow-16 6069 196905 ns/op 0 B/op 0 allocs/op
BenchmarkCachePutGetNoOverflow-16 1870 644236 ns/op 262 B/op 0 allocs/op
BenchmarkCachePutOverflow-16 2296 509279 ns/op 34 B/op 0 allocs/op
BenchmarkCachePutGetOverflow-16 1726 671510 ns/op 45 B/op 0 allocs/op
New cache implementation performance on 16 cores:
BenchmarkCachePutNoOverFlow-16 13549 82413 ns/op 0 B/op 0 allocs/op
BenchmarkCacheGetNoOverflow-16 30274 38997 ns/op 0 B/op 0 allocs/op
BenchmarkCachePutGetNoOverflow-16 8512 126239 ns/op 0 B/op 0 allocs/op
BenchmarkCachePutOverflow-16 13884 88124 ns/op 1 B/op 0 allocs/op
BenchmarkCachePutGetOverflow-16 7903 131299 ns/op 3 B/op 0 allocs/op
```
From the benchmarks above, we can see that the new implementation is ~5
times faster than the old one.
---------
Co-authored-by: f41gh7 <nik@victoriametrics.com>
This PR fixes#5796. See the points 6 and 7 in `Steps to reproduce`:
> Now let's set time to only 5ms past the timestamp of the first point,
since even 199ms worked for the second point. Surprise, the point isn't
returned 💥:
>
> ```curl -s $VMQURL -d 'query=series1' -d 'time=1707123456705' -d
'step=1ms' | grep 10 # nothing!```
>
> But, 4ms works: 🤨🤔
>
> ```curl -s $VMQURL -d 'query=series1' -d 'time=1707123456704' -d
'step=1ms' | grep 10 # found```
This happens so because the actual step becomes 5ms due to jitter being
applied. THe fix is to do not apply jitter if scrape interval was not
detected (the case when vmstorage returns only one result). In this case
the scrape interval is set to `5m+step`.
An integration test has been added to check the steps to reproduce and
then to confirm that fix works. Note that the cluster tests are
currently disabled because the fix is not in cluster branch yet.
The following checks are **mandatory**:
- [x] My change adheres [VictoriaMetrics contributing
guidelines](https://docs.victoriametrics.com/contributing/).
---------
Signed-off-by: Artem Fetishev <rtm@victoriametrics.com>
This is a follow-up after 3120dc2
- Consistently use key for rollupCache in multitenant mode cache keys use different authTokens. Previously it could lead to panic in rare cases when cache state was inconsistent.
- Do not share `err` variable across goroutines for `processBlock` function. It could lead to data races.
Related issue https://github.com/VictoriaMetrics/VictoriaMetrics/issues/7549
---------
Signed-off-by: Andrei Baidarov <abaidarov@yandex.ru>
Co-authored-by: f41gh7 <nik@victoriametrics.com>
Previously, for `^` aka pow function calls, VictoriaMetrics returned `1`
if left arg was Nan. For example, given query=`(hour()==2)^1` returns 1
for NaN produced by hour() == 2 function. It added additional non-exist
datapoints to the timeseries.
This commit port bugfix from `metricql` package and adds test for it.
Now, VictoriaMetrics
correctly returns `NaN` for such cases.
Related issue:
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/7359
Signed-off-by: f41gh7 <nik@victoriametrics.com>
(cherry picked from commit bb399518db)
This commit fixes panic for multitenant requests and empty storage node responses for tenants api.
It also optimizes `populateSqTenantTokensIfNeeded` function calls, by making it only once for query request. Previously it was incorrectly called multiple times per each storage node request.
Related issue:
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/7549
---------
Signed-off-by: f41gh7 <nik@victoriametrics.com>
Co-authored-by: f41gh7 <nik@victoriametrics.com>
### Describe Your Changes
I don't like this solution, but it works. Other possible solutions
described in an issue
fixes https://github.com/VictoriaMetrics/VictoriaMetrics/issues/7068
### Checklist
The following checks are **mandatory**:
- [ ] My change adheres [VictoriaMetrics contributing
guidelines](https://docs.victoriametrics.com/contributing/).
---------
Signed-off-by: hagen1778 <roman@victoriametrics.com>
Co-authored-by: hagen1778 <roman@victoriametrics.com>
(cherry picked from commit a88f896b43)
The purpose of this change is to reduce confusion between using
`flag.Duration` and `flagutils.Duration`. The reason is that
`flagutils.Duration` was mistakenly used for cases that required `m`
support. See
ab0d31a7b0
The change in name should clearly indicate the purpose of this data
type.
Please provide a brief description of the changes you made. Be as
specific as possible to help others understand the purpose and impact of
your modifications.
The following checks are **mandatory**:
- [ ] My change adheres [VictoriaMetrics contributing
guidelines](https://docs.victoriametrics.com/contributing/).
Signed-off-by: hagen1778 <roman@victoriametrics.com>
### Describe Your Changes
evalInstantRollup could have overreport the number of fetched series if
`offset` checks will result into retry. This change updates fetched
series only if these checks were successful.
It also adds a comment to another potential place of over-reporting
series fetched. It doesn't fix it, because it would require spending
extra resources on such a check, while discrepancy in seriesFetched
doesn't affect calculations in any way.
Probably related to
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/7170
### Checklist
The following checks are **mandatory**:
- [x] My change adheres [VictoriaMetrics contributing
guidelines](https://docs.victoriametrics.com/contributing/).
Signed-off-by: hagen1778 <roman@victoriametrics.com>
(cherry picked from commit ebd393d8b3)
Signed-off-by: hagen1778 <roman@victoriametrics.com>
### Describe Your Changes
Added an ability to query data across multiple tenants. See:
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1434
Currently, the following endpoints work with multi-tenancy:
- /prometheus/api/v1/query
- /prometheus/api/v1/query_range
- /prometheus/api/v1/series
- /prometheus/api/v1/labels
- /prometheus/api/v1/label/<label_name>/values
- /prometheus/api/v1/status/active_queries
- /prometheus/api/v1/status/top_queries
- /prometheus/api/v1/status/tsdb
- /prometheus/api/v1/export
- /prometheus/api/v1/export/csv
- /vmui
A note regarding VMUI: endpoints such as `active_queries` and
`top_queries` have been updated to indicate whether query was a
single-tenant or multi-tenant, but UI needs to be updated to display
this info.
cc: @Loori-R
---------
Signed-off-by: Zakhar Bessarab <z.bessarab@victoriametrics.com>
Signed-off-by: f41gh7 <nik@victoriametrics.com>
Co-authored-by: f41gh7 <nik@victoriametrics.com>
It is expected that range_first and range_last functions return non-nan const value across all the points
if the original series contains at least a single non-NaN value. Previously this rule was violated for NaN data points
in the original series. This could confuse users.
While at it, add tests for series with NaN values across all the range_* and running_* functions, in order to maintain
consistent handling of NaN values across these functions.
The %q formatter may result in incorrectly formatted JSON string if the original string
contains special chars such as \x1b . They must be encoded as \u001b , otherwise the resulting JSON string
cannot be parsed by JSON parsers.
This is a follow-up for c0caa69939
See https://github.com/VictoriaMetrics/victorialogs-datasource/issues/24
Consistently using t.Fatal* simplifies the test code and makes it less fragile, since it is common error
to forget to make proper cleanup after t.Error* call. Also t.Error* calls do not provide any practical
benefits when some tests fail. They just clutter test output with additional noise information,
which do not help in fixing failing tests most of the time.
This is a follow-up for a9525da8a4
### Describe Your Changes
In most cases histograms are exposed in sorted manner with lower buckets
being first. This means that during scraping buckets with lower bounds
have higher chance of being updated earlier than upper ones.
Previously, values were propagated from upper to lower bounds, which
means that in most cases that would produce results higher than expected
once all buckets will become updated.
Propagating from upper bound effectively limits highest value of
histogram to the value of previous scrape. Once the data will become
consistent in the subsequent evaluation this causes spikes in the
result.
Changing propagation to be from lower to higher buckets reduces value
spikes in most cases due to nature of the original inconsistency.
See: https://github.com/VictoriaMetrics/VictoriaMetrics/issues/4580
An example histogram with previous(red) and updated(blue) versions:
This also makes logic of filling nan values with lower buckets values: [1 2 3 nan nan nan] => [1 2 3 3 3 3] obsolete.
Since buckets are now fixed from lower ones to upper this happens in the main loop, so there is no need in a second one.
---------
Signed-off-by: Zakhar Bessarab <z.bessarab@victoriametrics.com>
Signed-off-by: hagen1778 <roman@victoriametrics.com>
Co-authored-by: Andrii Chubatiuk <andrew.chubatiuk@gmail.com>
Co-authored-by: hagen1778 <roman@victoriametrics.com>
(cherry picked from commit 6a4bd5049b)
'any' type is supported starting from Go1.18. Let's consistently use it
instead of 'interface{}' type across the code base, since `any` is easier to read than 'interface{}'.
Use metricsql.IsLikelyInvalid() function for determining whether the given query is likely invalid,
e.g. there is high change the query is incorrectly written, so it will return unexpected results.
The query is invalid most of the time if it passes something other than series selector into rollup function.
For example:
- rate(sum(foo))
- rate(foo + bar)
- rate(foo > bar)
Improtant note: the query is considered valid if it misses the lookbehind window in square brackes inside rollup function,
e.g. rate(foo), since this is very convenient MetricsQL extention to PromQL, and this query returns the expected results
most of the time.
Other unsafe query types can be added in the future into metricsql.IsLikelyInvalid().
TODO: probably, the -search.disableImplicitConversion command-line flag must be set by default in the future releases of VictoriaMetrics.
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/4338
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/pull/6180
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/pull/6450
Check for ranged vector arguments in aggregate expressions when
`-search.disableImplicitConversion` or `-search.logImplicitConversion`
are enabled.
For example, `sum(up[5m])` will fail to execute if these flags are set.
### Describe Your Changes
Please provide a brief description of the changes you made. Be as
specific as possible to help others understand the purpose and impact of
your modifications.
### Checklist
The following checks are **mandatory**:
- [*] My change adheres [VictoriaMetrics contributing
guidelines](https://docs.victoriametrics.com/contributing/).
---------
Signed-off-by: hagen1778 <roman@victoriametrics.com>
(cherry picked from commit 6149adbe10)
* metricsql: fix label_join() when `dst_label` is equal to one of the `src_label`
* Update app/vmselect/promql/transform.go
* Update docs/CHANGELOG.md
---------
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
Prevsiously they were swapped - the first arg should be the label name and the second arg should be label filters
This is a follow-up for e389b7b959e8144fdff5075bf7a5a39b2b0c6dd3
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/5847
This should improve maintainability of the code related to rollup functions,
since it is located in rollup.go
While at it, properly return empty results from holt_winters(), rate_over_sum(),
sum2_over_time(), geomean_over_time() and distinct_over_time() when there are no real samples
on the selected lookbehind window. Previously the previous sample value was mistakenly
returned from these functions.
changes(), increases_over_time() and resets() shouldn't take into account
value changes, which may occur because of precision errors.
The maximum guaranteed precision for raw samples stored in VictoriaMetrics is 12 decimal digits.
So do not count relative changes for values if they are smaller than 1e-12 comparing to the value.
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/767
* app/vmselect: set proper timestamp for cached instant responses
The change updates `getSumInstantValues` to prefer timestamp
from the most recent results. Before, timestamp from cached series
was used.
The old behavior had negative impact on recording rules as they
were getting responses with shifted timestamps in past.
Subsequent recording or alerting rules fetching results of these
recording rules could get no result due to staleness interval.
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/5659
Signed-off-by: hagen1778 <roman@victoriametrics.com>
* wip
---------
Signed-off-by: hagen1778 <roman@victoriametrics.com>
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
