When Prometheus restarts it creates every series read in from the WAL,
but many of those series will be finished, and never receive any more
samples. By defering allocation of the txRing slice to when it is first
needed, we save 32 bytes per stale series.
Signed-off-by: Bryan Boreham <bjboreham@gmail.com>
When reading the WAL this method is called with buffers from a pool, on
multiple goroutines. Pre-allocating sufficient size avoids slow growth
and many reallocations in `append`.
Signed-off-by: Bryan Boreham <bjboreham@gmail.com>
On a 32 bit architecture the size of int is 32 bits. Thus converting from
int64, uint64 can overflow it and flip the sign.
Try for yourself in playground:
package main
import "fmt"
func main() {
x := int64(0x1F0000001)
y := int64(1)
z := int32(x - y) // numerically this is 0x1F0000000
fmt.Printf("%v\n", z)
}
Prints -268435456 as if x was smaller.
Followup to #12650
Signed-off-by: György Krajcsovits <gyorgy.krajcsovits@grafana.com>
* Additionally wrap WBL replay error
Although WBL replay is already wrapped with errLoadWbl,
there are other errors that can happen during a WBL replay.
We should not try to repair WAL in those cases.
This commit additionally wraps the final error in Head.Init again
with errLoadWbl so that WBL replay errors can be identified properly.
Signed-off-by: Ganesh Vernekar <ganeshvern@gmail.com>
Signed-off-by: Jesus Vazquez <jesusvzpg@gmail.com>
Co-authored-by: Jesus Vazquez <jesusvzpg@gmail.com>
Signed-off-by: Levi Harrison <git@leviharrison.dev>
* Additionally wrap WBL replay error
Although WBL replay is already wrapped with errLoadWbl,
there are other errors that can happen during a WBL replay.
We should not try to repair WAL in those cases.
This commit additionally wraps the final error in Head.Init again
with errLoadWbl so that WBL replay errors can be identified properly.
Signed-off-by: Ganesh Vernekar <ganeshvern@gmail.com>
Signed-off-by: Jesus Vazquez <jesusvzpg@gmail.com>
Co-authored-by: Jesus Vazquez <jesusvzpg@gmail.com>
Reverts change from https://github.com/prometheus/prometheus/pull/12906
The benchmarks show that it's slower when intersecting, which is a
common usage for ListPostings (when intersecting matchers from Head)
(old is before #12906, new is #12906):
│ old │ new │
│ sec/op │ sec/op vs base │
Intersect/LongPostings1-16 20.54µ ± 1% 21.11µ ± 1% +2.76% (p=0.000 n=20)
Intersect/LongPostings2-16 51.03m ± 1% 52.40m ± 2% +2.69% (p=0.000 n=20)
Intersect/ManyPostings-16 194.2m ± 3% 332.1m ± 1% +71.00% (p=0.000 n=20)
geomean 5.882m 7.161m +21.74%
Signed-off-by: Oleg Zaytsev <mail@olegzaytsev.com>
It's implicit, but should be explicit. It is invalid to call At() after
a failed call to Next() or Seek().
Following up on https://github.com/prometheus/prometheus/pull/12906
Signed-off-by: Oleg Zaytsev <mail@olegzaytsev.com>
The Next() call of ListPostings() was updating two values, while we can
just update the position. This is up to 30% faster for high number of
Postings.
goos: linux
goarch: amd64
pkg: github.com/prometheus/prometheus/tsdb/index
cpu: 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz
│ old │ new │
│ sec/op │ sec/op vs base │
ListPostings/count=100-16 819.2n ± 0% 732.6n ± 0% -10.58% (p=0.000 n=20)
ListPostings/count=1000-16 2.685µ ± 1% 2.017µ ± 0% -24.88% (p=0.000 n=20)
ListPostings/count=10000-16 21.43µ ± 1% 14.81µ ± 0% -30.91% (p=0.000 n=20)
ListPostings/count=100000-16 209.4µ ± 1% 143.3µ ± 0% -31.55% (p=0.000 n=20)
ListPostings/count=1000000-16 2.086m ± 1% 1.436m ± 1% -31.18% (p=0.000 n=20)
geomean 29.02µ 21.41µ -26.22%
We're talking about microseconds here, but they just keep adding.
Signed-off-by: Oleg Zaytsev <mail@olegzaytsev.com>
This avoids situations where metrics are scraped before the data they
are trying to look at is initialized.
Signed-off-by: Bryan Boreham <bjboreham@gmail.com>
Currently, the two goroutines race against each other and it's possible that the main test goroutine finishes way earlier than appendSeries has had a chance to run at all.
I tested this change by breaking the code that X fixed and running the race test 100 times. Without the additional time.Sleep the test failed 11 times. With the sleep it failed 65 out of the 100 runs. Which is still not ideal, but it's a step forward.
Signed-off-by: Dimitar Dimitrov <dimitar.dimitrov@grafana.com>
The test was introduced in # but was changed during the code review and not reran with the faulty code since then.
Closes #
Signed-off-by: Dimitar Dimitrov <dimitar.dimitrov@grafana.com>
Before cutting a new XOR chunk in case the chunk goes over the size
limit, check that the timestamp is in order and not equal or older
than the latest sample in the old chunk.
Signed-off-by: György Krajcsovits <gyorgy.krajcsovits@grafana.com>
TestHeadDetectsDuplcateSampleAtSizeLimit tests a regression where a
duplicate sample,is appended to the head, right when the head chunk is
at the size limit. The test adds all samples as duplicate, thus
expecting that the result has exactly half of the samples.
Signed-off-by: György Krajcsovits <gyorgy.krajcsovits@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>
Case a) empty span is at the beginning of the spans.
Case b) two consequtive empty spans with positive offsets.
Signed-off-by: György Krajcsovits <gyorgy.krajcsovits@grafana.com>
* Benchmark WBL
Extended WAL benchmark test with WBL parts too - added basic cases for
OOO handling - a percentage of series have a percentage of samples set
as OOO ones.
Signed-off-by: Fiona Liao <fiona.y.liao@gmail.com>
* Fix handling of explicit counter reset header in histograms.
Explicit counter reset were being ignored.
Also there was no unit test coverage.
Add test case for the first sample in a chunk.
Add test case for non first sample in chunk.
Signed-off-by: György Krajcsovits <gyorgy.krajcsovits@grafana.com>
---------
Signed-off-by: György Krajcsovits <gyorgy.krajcsovits@grafana.com>
Otherwise we have a highly unusual situation of over 100 chunks
in the headChunks list of each series, which heavily skews
performance.
Signed-off-by: Bryan Boreham <bjboreham@gmail.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>
`rand.Read` has been deprecated since Go 1.20
`crypto/rand.Read` is more appropriate
Ref: https://tip.golang.org/doc/go1.20
Signed-off-by: Michal Biesek <michalbiesek@gmail.com>
Add a chunk size limit in bytes
This creates a hard cap for XOR chunks of 1024 bytes.
The limit for histogram chunk is also 1024 bytes, but it is a soft limit as a histogram has a dynamic size, and even a single one could be larger than 1024 bytes.
This also avoids cutting new histogram chunks if the existing chunk has fewer than 10 histograms yet. In that way, we are accepting "jumbo chunks" in order to have at least 10 histograms in a chunk, allowing compression to kick in.
Signed-off-by: Justin Lei <justin.lei@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>
Simlar to cleanup of WAL files on startup, cleanup temporary
chunk_snapshot dirs. This prevents storage space leaks due to terminated
snapshots on shutdown.
Signed-off-by: SuperQ <superq@gmail.com>
When a particular SeriesLifecycleCallback tries to optimize and run
closer to the Head, keeping track of the HeadSeriesRef instead of the
labelsets, it's impossible to handle the PostDeletion callback properly
as there's no way to know which series refs were deleted from the head.
This changes the callback to provide the series refs alongside the
labelsets, so the implementation can choose what to do.
Signed-off-by: Oleg Zaytsev <mail@olegzaytsev.com>
The most common case is to have a nil error when appending series, so
let's check that first instead of checking the 3 error types first.
Signed-off-by: Oleg Zaytsev <mail@olegzaytsev.com>
Currently memSeries holds a single head chunk in-memory and a slice of mmapped chunks.
When append() is called on memSeries it might decide that a new headChunk is needed to use for given append() call.
If that happens it will first mmap existing head chunk and only after that happens it will create a new empty headChunk and continue appending
our sample to it.
Since appending samples uses write lock on memSeries no other read or write can happen until any append is completed.
When we have an append() that must create a new head chunk the whole memSeries is blocked until mmapping of existing head chunk finishes.
Mmapping itself uses a lock as it needs to be serialised, which means that the more chunks to mmap we have the longer each chunk might wait
for it to be mmapped.
If there's enough chunks that require mmapping some memSeries will be locked for long enough that it will start affecting
queries and scrapes.
Queries might timeout, since by default they have a 2 minute timeout set.
Scrapes will be blocked inside append() call, which means there will be a gap between samples. This will first affect range queries
or calls using rate() and such, since the time range requested in the query might have too few samples to calculate anything.
To avoid this we need to remove mmapping from append path, since mmapping is blocking.
But this means that when we cut a new head chunk we need to keep the old one around, so we can mmap it later.
This change makes memSeries.headChunk a linked list, memSeries.headChunk still points to the 'open' head chunk that receives new samples,
while older, yet to be mmapped, chunks are linked to it.
Mmapping is done on a schedule by iterating all memSeries one by one. Thanks to this we control when mmapping is done, since we trigger
it manually, which reduces the risk that it will have to compete for mmap locks with other chunks.
Signed-off-by: Łukasz Mierzwa <l.mierzwa@gmail.com>
Bryan Boreham