promql: Improve histogram_quantile calculation for classic buckets
Tiny differences between classic buckets are most likely caused by floating point precision issues. With this commit, relative changes below a certain threshold are ignored. This makes the result of histogram_quantile more meaningful, and also avoids triggering the _input to histogram_quantile needed to be fixed for monotonicity_ annotations in unactionable cases.
This commit also adds explanation of the new adjustment and of the monotonicity annotation to the documentation of `histogram_quantile`.
---------
Signed-off-by: Jeanette Tan <jeanette.tan@grafana.com>
Return annotations (warnings and infos) from PromQL queries
This generalizes the warnings we have already used before (but only for problems with remote read) as "annotations".
Annotations can be warnings or infos (the latter could be false positives). We do not treat them different in the API for now and return them all as "warnings". It would be easy to distinguish them and return infos separately, should that appear useful in the future.
The new annotations are then used to create a lot of warnings or infos during PromQL evaluations. Partially these are things we have wanted for a long time (e.g. inform the user that they have applied `rate` to a metric that doesn't look like a counter), but the new native histograms have created even more needs for those annotations (e.g. if a query tries to aggregate float numbers with histograms).
The annotations added here are not yet complete. A prominent example would be a warning about a range too short for a rate calculation. But such a warnings is more tricky to create with good fidelity and we will tackle it later.
Another TODO is to take annotations into account when evaluating recording rules.
---------
Signed-off-by: Jeanette Tan <jeanette.tan@grafana.com>
promql: Extend testing framework to support native histograms
This includes both the internal testing framework as well as the rules unit test feature of promtool.
This also adds a bunch of basic tests. Many of the code level tests can now be converted to tests within the framework, and more tests can be added easily.
---------
Signed-off-by: Harold Dost <h.dost@criteo.com>
Signed-off-by: Gregor Zeitlinger <gregor.zeitlinger@grafana.com>
Signed-off-by: Stephen Lang <stephen.lang@grafana.com>
Co-authored-by: Harold Dost <h.dost@criteo.com>
Co-authored-by: Stephen Lang <stephen.lang@grafana.com>
Co-authored-by: Gregor Zeitlinger <gregor.zeitlinger@grafana.com>
So far, `ValidateHistogram` would not detect if the count did not
include the count in the zero bucket. This commit fixes the problem
and updates all the tests that have been undetected offenders so far.
Note that this problem would only ever create false negatives, so we
never falsely rejected to store a histogram because of it.
On the other hand, `ValidateFloatHistogram` has been to strict with
the count being at least as large as the sum of the counts in all the
buckets. Float precision issues could create false positives here, see
products of PromQL evaluations, it's actually quite hard to put an
upper limit no the floating point imprecision. Users could produce the
weirdest expressions, maxing out float precision problems. Therefore,
this commit simply removes that particular check from
`ValidateFloatHistogram`.
Signed-off-by: beorn7 <beorn@grafana.com>
Convert QueryOpts to an interface so that downstream projects like
https://github.com/thanos-community/promql-engine could extend the query
options with engine specific options that are not in the original
engine.
Will be used to enable query analysis per-query.
Signed-off-by: Giedrius Statkevičius <giedrius.statkevicius@vinted.com>
Handle more arithmetic operators and aggregators for native histograms
This includes operators for multiplication (formerly known as scaling), division, and subtraction. Plus aggregations for average and the avg_over_time function.
Stdvar and stddev will (for now) ignore histograms properly (rather than counting them but adding a 0 for them).
Signed-off-by: Jeanette Tan <jeanette.tan@grafana.com>
In other words: Instead of having a “polymorphous” `Point` that can
either contain a float value or a histogram value, use an `FPoint` for
floats and an `HPoint` for histograms.
This seemingly small change has a _lot_ of repercussions throughout
the codebase.
The idea here is to avoid the increase in size of `Point` arrays that
happened after native histograms had been added.
The higher-level data structures (`Sample`, `Series`, etc.) are still
“polymorphous”. The same idea could be applied to them, but at each
step the trade-offs needed to be evaluated.
The idea with this change is to do the minimum necessary to get back
to pre-histogram performance for functions that do not touch
histograms. Here are comparisons for the `changes` function. The test
data doesn't include histograms yet. Ideally, there would be no change
in the benchmark result at all.
First runtime v2.39 compared to directly prior to this commit:
```
name old time/op new time/op delta
RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 542µs ± 1% +38.58% (p=0.000 n=9+8)
RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 617µs ± 2% +36.48% (p=0.000 n=10+10)
RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.36ms ± 2% +21.58% (p=0.000 n=8+10)
RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 8.94ms ± 1% +14.21% (p=0.000 n=10+10)
RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.30ms ± 1% +10.67% (p=0.000 n=9+10)
RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.10ms ± 1% +11.82% (p=0.000 n=10+10)
RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 11.8ms ± 1% +12.50% (p=0.000 n=8+10)
RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 87.4ms ± 1% +12.63% (p=0.000 n=9+9)
RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 32.8ms ± 1% +8.01% (p=0.000 n=10+10)
RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.6ms ± 2% +9.64% (p=0.000 n=10+10)
RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 117ms ± 1% +11.69% (p=0.000 n=10+10)
RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 876ms ± 1% +11.83% (p=0.000 n=9+10)
```
And then runtime v2.39 compared to after this commit:
```
name old time/op new time/op delta
RangeQuery/expr=changes(a_one[1d]),steps=1-16 391µs ± 2% 547µs ± 1% +39.84% (p=0.000 n=9+8)
RangeQuery/expr=changes(a_one[1d]),steps=10-16 452µs ± 2% 616µs ± 2% +36.15% (p=0.000 n=10+10)
RangeQuery/expr=changes(a_one[1d]),steps=100-16 1.12ms ± 1% 1.26ms ± 1% +12.20% (p=0.000 n=8+10)
RangeQuery/expr=changes(a_one[1d]),steps=1000-16 7.83ms ± 1% 7.95ms ± 1% +1.59% (p=0.000 n=10+8)
RangeQuery/expr=changes(a_ten[1d]),steps=1-16 2.98ms ± 0% 3.38ms ± 2% +13.49% (p=0.000 n=9+10)
RangeQuery/expr=changes(a_ten[1d]),steps=10-16 3.66ms ± 1% 4.02ms ± 1% +9.80% (p=0.000 n=10+9)
RangeQuery/expr=changes(a_ten[1d]),steps=100-16 10.5ms ± 0% 10.8ms ± 1% +3.08% (p=0.000 n=8+10)
RangeQuery/expr=changes(a_ten[1d]),steps=1000-16 77.6ms ± 1% 78.1ms ± 1% +0.58% (p=0.035 n=9+10)
RangeQuery/expr=changes(a_hundred[1d]),steps=1-16 30.4ms ± 2% 33.5ms ± 4% +10.18% (p=0.000 n=10+10)
RangeQuery/expr=changes(a_hundred[1d]),steps=10-16 37.1ms ± 2% 40.0ms ± 1% +7.98% (p=0.000 n=10+10)
RangeQuery/expr=changes(a_hundred[1d]),steps=100-16 105ms ± 1% 107ms ± 1% +1.92% (p=0.000 n=10+10)
RangeQuery/expr=changes(a_hundred[1d]),steps=1000-16 783ms ± 3% 775ms ± 1% -1.02% (p=0.019 n=9+9)
```
In summary, the runtime doesn't really improve with this change for
queries with just a few steps. For queries with many steps, this
commit essentially reinstates the old performance. This is good
because the many-step queries are the one that matter most (longest
absolute runtime).
In terms of allocations, though, this commit doesn't make a dent at
all (numbers not shown). The reason is that most of the allocations
happen in the sampleRingIterator (in the storage package), which has
to be addressed in a separate commit.
Signed-off-by: beorn7 <beorn@grafana.com>
This is a bit more conservative than we could be. As long as a chunk
isn't the first in a block, we can be pretty sure that the previous
chunk won't disappear. However, the incremental gain of returning
NotCounterReset in these cases is probably very small and might not be
worth the code complications.
Wwith this, we now also pay attention to an explicitly set counter
reset during ingestion. While the case doesn't show up in practice
yet, there could be scenarios where the metric source knows there was
a counter reset even if it might not be visible from the values in the
histogram. It is also useful for testing.
Signed-off-by: beorn7 <beorn@grafana.com>
Extends Appender.AppendHistogram function to accept the FloatHistogram. TSDB supports appending, querying, WAL replay, for this new type of histogram.
Signed-off-by: Marc Tudurí <marctc@protonmail.com>
Signed-off-by: Ganesh Vernekar <ganeshvern@gmail.com>
Co-authored-by: Ganesh Vernekar <ganeshvern@gmail.com>
If we are populating series for a subquery then set the interval
parameter accordingly so that downstream users could use that
information.
Signed-off-by: Giedrius Statkevičius <giedrius.statkevicius@vinted.com>
And use the new method to call to compact Histograms during
parsing. This happens for both `Histogram` and `FloatHistogram`. In
this way, if targets decide to optimize the exposition size by merging
spans with empty buckets in between, we still get a normalized
results. It will also normalize away any valid but weird
representations like empty spans, spans with offset zero, and empty
buckets at the start or end of a span.
The implementation seemed easy at first as it just turns the
`compactBuckets` helper into a generic function (which now got its own
file). However, the integer Histograms have delta buckets instead of
absolute buckets, which had to be treated specially in the generic
`compactBuckets` function. To make sure it works, I have added plenty
of explicit tests for `Histogram` in addition to the `FloatHistogram`
tests.
I have also updated the doc comment for the `Compact` method.
Based on the insights now expressed in the doc comment, compacting
with a maxEmptyBuckets > 0 is rarely useful. Therefore, this commit
also sets the value to 0 in the two cases we were using 3 so far. We
might still want to reconsider, so I don't want to remove the
maxEmptyBuckets parameter right now.
Signed-off-by: beorn7 <beorn@grafana.com>
And a few cases of `EmptyLabels()`.
Replacing code which assumes the internal structure of `Labels`.
Signed-off-by: Bryan Boreham <bjboreham@gmail.com>
* Add histogram validation
Signed-off-by: Levi Harrison <git@leviharrison.dev>
* Correct negative offset validation
Signed-off-by: Levi Harrison <git@leviharrison.dev>
* Address review comments
Signed-off-by: Levi Harrison <git@leviharrison.dev>
* Validation benchmark
Signed-off-by: Levi Harrison <git@leviharrison.dev>
* Add more checks
Signed-off-by: Levi Harrison <git@leviharrison.dev>
* Attempt to fix tests
Signed-off-by: Levi Harrison <git@leviharrison.dev>
* Fix stuff
Signed-off-by: Levi Harrison <git@leviharrison.dev>
We print the stacktrace of a panic when query causes one, but there's no
information about the query itself, which makes it harder to debug and
reproduce the issue.
This adds the 'expr' string to the logged panic.
Signed-off-by: Łukasz Mierzwa <l.mierzwa@gmail.com>
This follow a simple function-based approach to access the count and
sum fields of a native Histogram. It might be more elegant to
implement “accessors” via the dot operator, as considered in the
brainstorming doc [1]. However, that would require the introduction of
a whole new concept in PromQL. For the PoC, we should be fine with the
function-based approch. Even the obvious inefficiencies (rate'ing a
whole histogram twice when we only want to rate each the count and the
sum once) could be optimized behind the scenes.
Note that the function-based approach elegantly solves the problem of
detecting counter resets in the sum of observations in the case of
negative observations. (Since the whole native Histogram is rate'd,
the counter reset is detected for the Histogram as a whole.)
We will decide later if an “accessor” approach is really needed. It
would change the example expression for average duration in
functions.md from
histogram_sum(rate(http_request_duration_seconds[10m]))
/
histogram_count(rate(http_request_duration_seconds[10m]))
to
rate(http_request_duration_seconds.sum[10m])
/
rate(http_request_duration_seconds.count[10m])
[1]: https://docs.google.com/document/d/1ch6ru8GKg03N02jRjYriurt-CZqUVY09evPg6yKTA1s/edit
Signed-off-by: beorn7 <beorn@grafana.com>
Essentially, this mirrors the existing behavior for negative buckets:
If a histogram has only negative buckets, the upper bound of the zero
bucket is assumed to be zero.
Furthermore, it makes sure that the zero bucket boundaries are not
modified if a histogram that has no buckets at all but samples in the
zero bucket.
Also, add an TODO to vet if we really want this behavior.
Signed-off-by: beorn7 <beorn@grafana.com>
* refactor: move from io/ioutil to io and os packages
* use fs.DirEntry instead of os.FileInfo after os.ReadDir
Signed-off-by: MOREL Matthieu <matthieu.morel@cnp.fr>
This exactly corresponds to the statistic compared against MaxSamples
during the course of query execution, so users can see how close their
queries are to a limit.
Co-authored-by: Harkishen Singh <harkishensingh@hotmail.com>
Co-authored-by: Andrew Bloomgarden <blmgrdn@amazon.com>
Signed-off-by: Andrew Bloomgarden <blmgrdn@amazon.com>
We always track total samples queried and add those to the standard set
of stats queries can report.
We also allow optionally tracking per-step samples queried. This must be
enabled both at the engine and query level to be tracked and rendered.
The engine flag is exposed via a Prometheus feature flag, while the
query flag is set when stats=all.
Co-authored-by: Alan Protasio <approtas@amazon.com>
Co-authored-by: Andrew Bloomgarden <blmgrdn@amazon.com>
Co-authored-by: Harkishen Singh <harkishensingh@hotmail.com>
Signed-off-by: Andrew Bloomgarden <blmgrdn@amazon.com>
Matthieu MOREL