While investigating lock contention on `MemPostings`, we saw that lots
of locking is happening in `LabelValues` and
`PostingsForLabelsMatching`, both copying the label values slices while
holding the mutex.
This adds an extra map that holds an append-only label values slice for
each one of the label names. Since the slice is append-only, it can be
copied without holding the mutex.
Signed-off-by: Oleg Zaytsev <mail@olegzaytsev.com>
Same as #15427 but for the new method added in #14144
Instead of allocating each ListPostings one by one, allocate them all in
one go.
Signed-off-by: Oleg Zaytsev <mail@olegzaytsev.com>
Instead of allocating ListPostings pointers one by one, allocate a slice
and take pointers from that. It's faster, and also generates less
garbage (NewListPostings is one of the top offenders in number of
allocations).
Signed-off-by: Oleg Zaytsev <mail@olegzaytsev.com>
Simple follow-up to #13620. Modify `tsdb.PostingsForMatchers` to use the optimized tsdb.IndexReader.PostingsForLabelMatching method also for inverse matching.
Introduce method `PostingsForAllLabelValues`, to avoid changing the existing method.
The performance is much improved for a subset of the cases; there are up to
~60% CPU gains and ~12.5% reduction in memory usage.
Remove `TestReader_InversePostingsForMatcherHonorsContextCancel` since
`inversePostingsForMatcher` only passes `ctx` to `IndexReader` implementations now.
Signed-off-by: Arve Knudsen <arve.knudsen@gmail.com>
This introduces back some unlocking that was removed in #13286 but in a
more balanced way, as suggested by @pracucci.
For TSDBs with a lot of churn, Delete() can take a couple of seconds,
and while it's holding the mutex, reads and writes are blocked waiting
for that mutex, increasing the number of connections handled and memory
usage.
This implementation pauses every 4K labels processed (note that also
compared to #13286 we're not processing all the label-values anymore,
but only the affected ones, because of #14307), makes sure that it's
possible to get the read lock, and waits for a few milliseconds more.
Signed-off-by: Oleg Zaytsev <mail@olegzaytsev.com>
Co-authored-by: Marco Pracucci <marco@pracucci.com>
This reverts commit 50ef0dc954.
Memory allocation goes so high in Prombench that the system is unusable.
Signed-off-by: Bryan Boreham <bjboreham@gmail.com>
* Tests for Mempostings.{Add,Get} data race
* Fix MemPostings.{Add,Get} data race
We can't modify the postings list that are held in MemPostings as they
might already be in use by some readers.
* Modify BenchmarkHeadStripeSeriesCreate to have common labels
If there are no common labels on the series, we don't excercise the
ordering part of MemSeries, as we're just creating slices of one element
for each label value.
---------
Signed-off-by: Oleg Zaytsev <mail@olegzaytsev.com>
We are still seeing lock contention on MemPostings.mtx, and MemPostings.Delete() is by far the most expensive operation on that mutex.
This adds parallelism to that method, trying to reduce the amount of time we spend with the mutex held.
Signed-off-by: Oleg Zaytsev <mail@olegzaytsev.com>
Go's built-in append() grows larger slices with factor 1.3, which means we do a lot more allocating and copying for larger postings.
Signed-off-by: Bryan Boreham <bjboreham@gmail.com>
* Pass affected labels to MemPostings.Delete
As suggested by @bboreham, we can track the labels of the deleted series
and avoid iterating through all the label/value combinations.
This looks much faster on the MemPostings.Delete call. We don't have a
benchmark on stripeSeries.gc() where we'll pay the price of iterating
the labels of each one of the deleted series.
Signed-off-by: Oleg Zaytsev <mail@olegzaytsev.com>
* MemPostings.PostingsForLabelMatching: let mutex go
This changes the `MemPostings.PostingsForLabelMatching` implementation
to stop holding the read mutex while matching the label values.
We've seen that this method can be slow when the matcher is expensive,
that's why we even added a context expiration check.
However, there are critical process that might be waiting on this mutex:
writes (adding new series) and compaction (deleting the
garbage-collected ones), so we should avoid holding it for a long period
of time.
Given that we've copied the values to a slice anyway, there's no need to
hold the lock while matching.
Signed-off-by: Oleg Zaytsev <mail@olegzaytsev.com>
* MemPostings: reduce locking/unlocking
MemPostings.Delete is called from Head.gc(), i.e. it gets the IDs of the
series that have churned.
I'd assume that many label values aren't affected by that churn at all,
so it doesn't make sense to touch the lock while checking them.
Signed-off-by: Oleg Zaytsev <mail@olegzaytsev.com>
Followup to #14096
Unfortunately the previous PR introduced this bug by not releasing the
lock before returning.
Signed-off-by: György Krajcsovits <gyorgy.krajcsovits@grafana.com>
* tsdb: check for context cancel before regex matching postings
Regex matching can be heavy if the regex takes a lot of cycles to
evaluate and we can get stuck evaluating postings for a long time
without this fix. The constant checkContextEveryNIterations=100
may be changed later.
Signed-off-by: György Krajcsovits <gyorgy.krajcsovits@grafana.com>
Add method `PostingsForLabelMatching` to `tsdb.IndexReader`, to obtain postings for labels with a certain name and values accepted by a provided callback, and use it from `tsdb.PostingsForMatchers`.
The intention is to optimize regexp matcher paths, especially not having to load all label values before matching on them.
Plus tests, and refactor some `tsdb/index.Reader` methods.
Benchmarking shows memory reduction up to ~100%, and speedup of up to ~50%.
Signed-off-by: Arve Knudsen <arve.knudsen@gmail.com>
Co-authored-by: Bartlomiej Plotka <bwplotka@gmail.com>
It's faster.
Note change to test - instead of requiring that the data structure is
identical to `EmptyPostings()`, check that calling `Next()` returns
false, which implies it was empty.
Also the check for context cancellation during initialization was
removed. Initialization should be a small portion of the work done
during merge, so it's not worth plumbing a context argument through.
Signed-off-by: Bryan Boreham <bjboreham@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>
Problem:
LabelValueStats - This will provide a list of the label names and memory used in bytes.
It is calculated by adding the length of all values for a given label name.
But internally Prometheus stores the name and the value independently for each series.
Solution:
MemPostings struct maintains the values to seriesRef map which is used
to get the number of series which contains the label values.
Using that LabelValueStats is calculated as: seriesCnt * len(value
name)
Signed-off-by: Baskar Shanmugam <baskar.shanmugam.career@gmail.com>
Wiser coders than myself have come to the conclusion that a `switch`
statement is almost always superior to a statement that includes any
`else if`.
The exceptions that I have found in our codebase are just these two:
* The `if else` is followed by an additional statement before the next
condition (separated by a `;`).
* The whole thing is within a `for` loop and `break` statements are
used. In this case, using `switch` would require tagging the `for`
loop, which probably tips the balance.
Why are `switch` statements more readable?
For one, fewer curly braces. But more importantly, the conditions all
have the same alignment, so the whole thing follows the natural flow
of going down a list of conditions. With `else if`, in contrast, all
conditions but the first are "hidden" behind `} else if `, harder to
spot and (for no good reason) presented differently from the first
condition.
I'm sure the aforemention wise coders can list even more reasons.
In any case, I like it so much that I have found myself recommending
it in code reviews. I would like to make it a habit in our code base,
without making it a hard requirement that we would test on the CI. But
for that, there has to be a role model, so this commit eliminates all
`if else` occurrences, unless it is autogenerated code or fits one of
the exceptions above.
Signed-off-by: beorn7 <beorn@grafana.com>
Use new experimental package `golang.org/x/exp/slices`.
slices.Sort works on values that are directly comparable, like ints,
so avoids the overhad of an interface call to `.Less()`.
Left tests unchanged, because they don't need the speed and it may be
a cross-check that slices.Sort gives the same answer.
Signed-off-by: Bryan Boreham <bjboreham@gmail.com>
Use new experimental package `golang.org/x/exp/slices`.
Some of the speedup comes from comparing SeriesRef (which is an int64)
directly rather than through an interface `.Less()` call; some comes
from exp/slices using "pattern-defeating quicksort(pdqsort)".
Signed-off-by: Bryan Boreham <bjboreham@gmail.com>
Signed-off-by: Bryan Boreham <bjboreham@gmail.com>
* tsdb: avoid slice-to-interface allocation in EnsureOrder
This is pulling the `seriesRefSlice` out of the loop, so the compiler
doesn't allocate a new one on the heap every time.
Signed-off-by: Bryan Boreham <bjboreham@gmail.com>
* tsdb: use pointer type in Pool for EnsureOrder
As noted by staticcheck, Pool prefers the objects in the pool to have
pointer type. This is a little more fiddly to code, but avoids
allocation of a wrapper object every time a slice is put into the pool.
Removed a comment that said fixing this has a performance penalty: not
borne out by benchmarks.
Signed-off-by: Bryan Boreham <bjboreham@gmail.com>
Unexported postingsWithIndexHeap's methods that don't need to be
exported, and added detailed comments.
Signed-off-by: Oleg Zaytsev <mail@olegzaytsev.com>
See this comment for detailed explanation:
https://github.com/prometheus/prometheus/pull/9907#issuecomment-1002189932
TL;DR: if we don't call Pop() on the heap implementation, we don't need
to return our param as an `interface{}` so we save an allocation.
This would be popped for every label value, so it can be thousands of
saved allocations here (see benchmarks).
Signed-off-by: Oleg Zaytsev <mail@olegzaytsev.com>
This creates a new `model` directory and moves all data-model related
packages over there:
exemplar labels relabel rulefmt textparse timestamp value
All the others are more or less utilities and have been moved to `util`:
gate logging modetimevfs pool runtime
Signed-off-by: beorn7 <beorn@grafana.com>
* TSDB: demistify seriesRefs and ChunkRefs
The TSDB package contains many types of series and chunk references,
all shrouded in uint types. Often the same uint value may
actually mean one of different types, in non-obvious ways.
This PR aims to clarify the code and help navigating to relevant docs,
usage, etc much quicker.
Concretely:
* Use appropriately named types and document their semantics and
relations.
* Make multiplexing and demuxing of types explicit
(on the boundaries between concrete implementations and generic
interfaces).
* Casting between different types should be free. None of the changes
should have any impact on how the code runs.
TODO: Implement BlockSeriesRef where appropriate (for a future PR)
Signed-off-by: Dieter Plaetinck <dieter@grafana.com>
* feedback
Signed-off-by: Dieter Plaetinck <dieter@grafana.com>
* agent: demistify seriesRefs and ChunkRefs
Signed-off-by: Dieter Plaetinck <dieter@grafana.com>
This saves memory, effort and locking.
Since every symbol is also added to postings, `Symbols()` can be
implemented there instead. This now has to build a map for
deduplication, but `Symbols()` is only called for compaction, and `gc()`
used to rebuild the symbols map after every compaction so not an
additional cost.
Signed-off-by: Bryan Boreham <bjboreham@gmail.com>
