prometheus/storage/local/series.go

729 lines
23 KiB
Go
Raw Normal View History

// Copyright 2014 The Prometheus Authors
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
package local
import (
"sort"
"sync"
"time"
"github.com/prometheus/common/model"
"github.com/prometheus/prometheus/storage/metric"
)
const (
// chunkDescEvictionFactor is a factor used for chunkDesc eviction (as opposed
// to evictions of chunks, see method evictOlderThan. A chunk takes about 20x
// more memory than a chunkDesc. With a chunkDescEvictionFactor of 10, not more
// than a third of the total memory taken by a series will be used for
// chunkDescs.
chunkDescEvictionFactor = 10
headChunkTimeout = time.Hour // Close head chunk if not touched for that long.
)
// fingerprintSeriesPair pairs a fingerprint with a memorySeries pointer.
type fingerprintSeriesPair struct {
fp model.Fingerprint
series *memorySeries
}
// seriesMap maps fingerprints to memory series. All its methods are
// goroutine-safe. A SeriesMap is effectively is a goroutine-safe version of
// map[model.Fingerprint]*memorySeries.
type seriesMap struct {
mtx sync.RWMutex
m map[model.Fingerprint]*memorySeries
}
// newSeriesMap returns a newly allocated empty seriesMap. To create a seriesMap
// based on a prefilled map, use an explicit initializer.
func newSeriesMap() *seriesMap {
return &seriesMap{m: make(map[model.Fingerprint]*memorySeries)}
}
// length returns the number of mappings in the seriesMap.
func (sm *seriesMap) length() int {
sm.mtx.RLock()
defer sm.mtx.RUnlock()
return len(sm.m)
}
// get returns a memorySeries for a fingerprint. Return values have the same
// semantics as the native Go map.
func (sm *seriesMap) get(fp model.Fingerprint) (s *memorySeries, ok bool) {
sm.mtx.RLock()
defer sm.mtx.RUnlock()
s, ok = sm.m[fp]
return
}
// put adds a mapping to the seriesMap. It panics if s == nil.
func (sm *seriesMap) put(fp model.Fingerprint, s *memorySeries) {
sm.mtx.Lock()
defer sm.mtx.Unlock()
if s == nil {
panic("tried to add nil pointer to seriesMap")
}
sm.m[fp] = s
}
// del removes a mapping from the series Map.
func (sm *seriesMap) del(fp model.Fingerprint) {
sm.mtx.Lock()
defer sm.mtx.Unlock()
delete(sm.m, fp)
}
// iter returns a channel that produces all mappings in the seriesMap. The
// channel will be closed once all fingerprints have been received. Not
// consuming all fingerprints from the channel will leak a goroutine. The
// semantics of concurrent modification of seriesMap is the similar as the one
// for iterating over a map with a 'range' clause. However, if the next element
// in iteration order is removed after the current element has been received
// from the channel, it will still be produced by the channel.
func (sm *seriesMap) iter() <-chan fingerprintSeriesPair {
ch := make(chan fingerprintSeriesPair)
go func() {
sm.mtx.RLock()
for fp, s := range sm.m {
sm.mtx.RUnlock()
ch <- fingerprintSeriesPair{fp, s}
sm.mtx.RLock()
}
sm.mtx.RUnlock()
close(ch)
}()
return ch
}
// fpIter returns a channel that produces all fingerprints in the seriesMap. The
// channel will be closed once all fingerprints have been received. Not
// consuming all fingerprints from the channel will leak a goroutine. The
// semantics of concurrent modification of seriesMap is the similar as the one
// for iterating over a map with a 'range' clause. However, if the next element
// in iteration order is removed after the current element has been received
// from the channel, it will still be produced by the channel.
func (sm *seriesMap) fpIter() <-chan model.Fingerprint {
ch := make(chan model.Fingerprint)
go func() {
sm.mtx.RLock()
for fp := range sm.m {
sm.mtx.RUnlock()
ch <- fp
sm.mtx.RLock()
}
sm.mtx.RUnlock()
close(ch)
}()
return ch
}
type memorySeries struct {
metric model.Metric
// Sorted by start time, overlapping chunk ranges are forbidden.
chunkDescs []*chunkDesc
// The index (within chunkDescs above) of the first chunkDesc that
// points to a non-persisted chunk. If all chunks are persisted, then
// persistWatermark == len(chunkDescs).
persistWatermark int
// The modification time of the series file. The zero value of time.Time
// is used to mark an unknown modification time.
modTime time.Time
// The chunkDescs in memory might not have all the chunkDescs for the
// chunks that are persisted to disk. The missing chunkDescs are all
// contiguous and at the tail end. chunkDescsOffset is the index of the
// chunk on disk that corresponds to the first chunkDesc in memory. If
// it is 0, the chunkDescs are all loaded. A value of -1 denotes a
// special case: There are chunks on disk, but the offset to the
// chunkDescs in memory is unknown. Also, in this special case, there is
// no overlap between chunks on disk and chunks in memory (implying that
// upon first persisting of a chunk in memory, the offset has to be
// set).
chunkDescsOffset int
// The savedFirstTime field is used as a fallback when the
// chunkDescsOffset is not 0. It can be used to save the firstTime of the
// first chunk before its chunk desc is evicted. In doubt, this field is
// just set to the oldest possible timestamp.
savedFirstTime model.Time
// The timestamp of the last sample in this series. Needed for fast
// access for federation and to ensure timestamp monotonicity during
// ingestion.
lastTime model.Time
// The last ingested sample value. Needed for fast access for
// federation.
lastSampleValue model.SampleValue
// Whether lastSampleValue has been set already.
lastSampleValueSet bool
// Whether the current head chunk has already been finished. If true,
// the current head chunk must not be modified anymore.
headChunkClosed bool
// Whether the current head chunk is used by an iterator. In that case,
// a non-closed head chunk has to be cloned before more samples are
// appended.
headChunkUsedByIterator bool
// Whether the series is inconsistent with the last checkpoint in a way
2015-03-18 18:09:07 +00:00
// that would require a disk seek during crash recovery.
dirty bool
}
// newMemorySeries returns a pointer to a newly allocated memorySeries for the
// given metric. chunkDescs and modTime in the new series are set according to
// the provided parameters. chunkDescs can be nil or empty if this is a
// genuinely new time series (i.e. not one that is being unarchived). In that
// case, headChunkClosed is set to false, and firstTime and lastTime are both
// set to model.Earliest. The zero value for modTime can be used if the
// modification time of the series file is unknown (e.g. if this is a genuinely
// new series).
func newMemorySeries(m model.Metric, chunkDescs []*chunkDesc, modTime time.Time) (*memorySeries, error) {
var err error
firstTime := model.Earliest
lastTime := model.Earliest
if len(chunkDescs) > 0 {
firstTime = chunkDescs[0].firstTime()
if lastTime, err = chunkDescs[len(chunkDescs)-1].lastTime(); err != nil {
return nil, err
}
}
return &memorySeries{
metric: m,
chunkDescs: chunkDescs,
headChunkClosed: len(chunkDescs) > 0,
savedFirstTime: firstTime,
lastTime: lastTime,
persistWatermark: len(chunkDescs),
modTime: modTime,
}, nil
}
// add adds a sample pair to the series. It returns the number of newly
// completed chunks (which are now eligible for persistence).
//
// The caller must have locked the fingerprint of the series.
func (s *memorySeries) add(v model.SamplePair) (int, error) {
if len(s.chunkDescs) == 0 || s.headChunkClosed {
newHead := newChunkDesc(newChunk(), v.Timestamp)
s.chunkDescs = append(s.chunkDescs, newHead)
s.headChunkClosed = false
} else if s.headChunkUsedByIterator && s.head().refCount() > 1 {
// We only need to clone the head chunk if the current head
// chunk was used in an iterator at all and if the refCount is
// still greater than the 1 we always have because the head
// chunk is not yet persisted. The latter is just an
// approximation. We will still clone unnecessarily if an older
// iterator using a previous version of the head chunk is still
// around and keep the head chunk pinned. We needed to track
// pins by version of the head chunk, which is probably not
// worth the effort.
chunkOps.WithLabelValues(clone).Inc()
// No locking needed here because a non-persisted head chunk can
// not get evicted concurrently.
s.head().c = s.head().c.clone()
s.headChunkUsedByIterator = false
}
chunks, err := s.head().add(v)
if err != nil {
return 0, err
}
s.head().c = chunks[0]
for _, c := range chunks[1:] {
s.chunkDescs = append(s.chunkDescs, newChunkDesc(c, c.firstTime()))
}
// Populate lastTime of now-closed chunks.
for _, cd := range s.chunkDescs[len(s.chunkDescs)-len(chunks) : len(s.chunkDescs)-1] {
cd.maybePopulateLastTime()
}
s.lastTime = v.Timestamp
s.lastSampleValue = v.Value
s.lastSampleValueSet = true
return len(chunks) - 1, nil
}
// maybeCloseHeadChunk closes the head chunk if it has not been touched for the
2015-03-18 18:09:07 +00:00
// duration of headChunkTimeout. It returns whether the head chunk was closed.
// If the head chunk is already closed, the method is a no-op and returns false.
//
// The caller must have locked the fingerprint of the series.
func (s *memorySeries) maybeCloseHeadChunk() bool {
if s.headChunkClosed {
return false
}
if time.Now().Sub(s.lastTime.Time()) > headChunkTimeout {
s.headChunkClosed = true
// Since we cannot modify the head chunk from now on, we
// don't need to bother with cloning anymore.
s.headChunkUsedByIterator = false
s.head().maybePopulateLastTime()
return true
}
return false
}
// evictChunkDescs evicts chunkDescs if there are chunkDescEvictionFactor times
// more than non-evicted chunks. iOldestNotEvicted is the index within the
// current chunkDescs of the oldest chunk that is not evicted.
func (s *memorySeries) evictChunkDescs(iOldestNotEvicted int) {
lenToKeep := chunkDescEvictionFactor * (len(s.chunkDescs) - iOldestNotEvicted)
if lenToKeep < len(s.chunkDescs) {
s.savedFirstTime = s.firstTime()
lenEvicted := len(s.chunkDescs) - lenToKeep
s.chunkDescsOffset += lenEvicted
s.persistWatermark -= lenEvicted
chunkDescOps.WithLabelValues(evict).Add(float64(lenEvicted))
numMemChunkDescs.Sub(float64(lenEvicted))
s.chunkDescs = append(
make([]*chunkDesc, 0, lenToKeep),
s.chunkDescs[lenEvicted:]...,
)
s.dirty = true
}
}
// dropChunks removes chunkDescs older than t. The caller must have locked the
// fingerprint of the series.
func (s *memorySeries) dropChunks(t model.Time) error {
keepIdx := len(s.chunkDescs)
for i, cd := range s.chunkDescs {
lt, err := cd.lastTime()
if err != nil {
return err
}
if !lt.Before(t) {
keepIdx = i
break
}
}
2016-04-15 17:18:40 +00:00
if keepIdx == len(s.chunkDescs) && !s.headChunkClosed {
// Never drop an open head chunk.
keepIdx--
}
if keepIdx <= 0 {
// Nothing to drop.
return nil
}
s.chunkDescs = append(
make([]*chunkDesc, 0, len(s.chunkDescs)-keepIdx),
s.chunkDescs[keepIdx:]...,
)
s.persistWatermark -= keepIdx
if s.persistWatermark < 0 {
panic("dropped unpersisted chunks from memory")
}
if s.chunkDescsOffset != -1 {
s.chunkDescsOffset += keepIdx
}
2016-04-15 17:18:40 +00:00
numMemChunkDescs.Sub(float64(keepIdx))
s.dirty = true
return nil
}
// preloadChunks is an internal helper method.
2015-05-06 14:53:12 +00:00
func (s *memorySeries) preloadChunks(
indexes []int, fp model.Fingerprint, mss *MemorySeriesStorage,
) (SeriesIterator, error) {
loadIndexes := []int{}
pinnedChunkDescs := make([]*chunkDesc, 0, len(indexes))
for _, idx := range indexes {
cd := s.chunkDescs[idx]
pinnedChunkDescs = append(pinnedChunkDescs, cd)
cd.pin(mss.evictRequests) // Have to pin everything first to prevent immediate eviction on chunk loading.
if cd.isEvicted() {
loadIndexes = append(loadIndexes, idx)
}
}
chunkOps.WithLabelValues(pin).Add(float64(len(pinnedChunkDescs)))
if len(loadIndexes) > 0 {
if s.chunkDescsOffset == -1 {
panic("requested loading chunks from persistence in a situation where we must not have persisted data for chunk descriptors in memory")
}
chunks, err := mss.loadChunks(fp, loadIndexes, s.chunkDescsOffset)
if err != nil {
// Unpin the chunks since we won't return them as pinned chunks now.
for _, cd := range pinnedChunkDescs {
cd.unpin(mss.evictRequests)
}
chunkOps.WithLabelValues(unpin).Add(float64(len(pinnedChunkDescs)))
return nopIter, err
}
for i, c := range chunks {
s.chunkDescs[loadIndexes[i]].setChunk(c)
}
}
Streamline series iterator creation This will fix issue #1035 and will also help to make issue #1264 less bad. The fundamental problem in the current code: In the preload phase, we quite accurately determine which chunks will be used for the query being executed. However, in the subsequent step of creating series iterators, the created iterators are referencing _all_ in-memory chunks in their series, even the un-pinned ones. In iterator creation, we copy a pointer to each in-memory chunk of a series into the iterator. While this creates a certain amount of allocation churn, the worst thing about it is that copying the chunk pointer out of the chunkDesc requires a mutex acquisition. (Remember that the iterator will also reference un-pinned chunks, so we need to acquire the mutex to protect against concurrent eviction.) The worst case happens if a series doesn't even contain any relevant samples for the query time range. We notice that during preloading but then we will still create a series iterator for it. But even for series that do contain relevant samples, the overhead is quite bad for instant queries that retrieve a single sample from each series, but still go through all the effort of series iterator creation. All of that is particularly bad if a series has many in-memory chunks. This commit addresses the problem from two sides: First, it merges preloading and iterator creation into one step, i.e. the preload call returns an iterator for exactly the preloaded chunks. Second, the required mutex acquisition in chunkDesc has been greatly reduced. That was enabled by a side effect of the first step, which is that the iterator is only referencing pinned chunks, so there is no risk of concurrent eviction anymore, and chunks can be accessed without mutex acquisition. To simplify the code changes for the above, the long-planned change of ValueAtTime to ValueAtOrBefore time was performed at the same time. (It should have been done first, but it kind of accidentally happened while I was in the middle of writing the series iterator changes. Sorry for that.) So far, we actively filtered the up to two values that were returned by ValueAtTime, i.e. we invested work to retrieve up to two values, and then we invested more work to throw one of them away. The SeriesIterator.BoundaryValues method can be removed once #1401 is fixed. But I really didn't want to load even more changes into this PR. Benchmarks: The BenchmarkFuzz.* benchmarks run 83% faster (i.e. about six times faster) and allocate 95% fewer bytes. The reason for that is that the benchmark reads one sample after another from the time series and creates a new series iterator for each sample read. To find out how much these improvements matter in practice, I have mirrored a beefy Prometheus server at SoundCloud that suffers from both issues #1035 and #1264. To reach steady state that would be comparable, the server needs to run for 15d. So far, it has run for 1d. The test server currently has only half as many memory time series and 60% of the memory chunks the main server has. The 90th percentile rule evaluation cycle time is ~11s on the main server and only ~3s on the test server. However, these numbers might get much closer over time. In addition to performance improvements, this commit removes about 150 LOC.
2016-02-16 17:47:50 +00:00
if !s.headChunkClosed && indexes[len(indexes)-1] == len(s.chunkDescs)-1 {
s.headChunkUsedByIterator = true
}
curriedQuarantineSeries := func(err error) {
mss.quarantineSeries(fp, s.metric, err)
}
Streamline series iterator creation This will fix issue #1035 and will also help to make issue #1264 less bad. The fundamental problem in the current code: In the preload phase, we quite accurately determine which chunks will be used for the query being executed. However, in the subsequent step of creating series iterators, the created iterators are referencing _all_ in-memory chunks in their series, even the un-pinned ones. In iterator creation, we copy a pointer to each in-memory chunk of a series into the iterator. While this creates a certain amount of allocation churn, the worst thing about it is that copying the chunk pointer out of the chunkDesc requires a mutex acquisition. (Remember that the iterator will also reference un-pinned chunks, so we need to acquire the mutex to protect against concurrent eviction.) The worst case happens if a series doesn't even contain any relevant samples for the query time range. We notice that during preloading but then we will still create a series iterator for it. But even for series that do contain relevant samples, the overhead is quite bad for instant queries that retrieve a single sample from each series, but still go through all the effort of series iterator creation. All of that is particularly bad if a series has many in-memory chunks. This commit addresses the problem from two sides: First, it merges preloading and iterator creation into one step, i.e. the preload call returns an iterator for exactly the preloaded chunks. Second, the required mutex acquisition in chunkDesc has been greatly reduced. That was enabled by a side effect of the first step, which is that the iterator is only referencing pinned chunks, so there is no risk of concurrent eviction anymore, and chunks can be accessed without mutex acquisition. To simplify the code changes for the above, the long-planned change of ValueAtTime to ValueAtOrBefore time was performed at the same time. (It should have been done first, but it kind of accidentally happened while I was in the middle of writing the series iterator changes. Sorry for that.) So far, we actively filtered the up to two values that were returned by ValueAtTime, i.e. we invested work to retrieve up to two values, and then we invested more work to throw one of them away. The SeriesIterator.BoundaryValues method can be removed once #1401 is fixed. But I really didn't want to load even more changes into this PR. Benchmarks: The BenchmarkFuzz.* benchmarks run 83% faster (i.e. about six times faster) and allocate 95% fewer bytes. The reason for that is that the benchmark reads one sample after another from the time series and creates a new series iterator for each sample read. To find out how much these improvements matter in practice, I have mirrored a beefy Prometheus server at SoundCloud that suffers from both issues #1035 and #1264. To reach steady state that would be comparable, the server needs to run for 15d. So far, it has run for 1d. The test server currently has only half as many memory time series and 60% of the memory chunks the main server has. The 90th percentile rule evaluation cycle time is ~11s on the main server and only ~3s on the test server. However, these numbers might get much closer over time. In addition to performance improvements, this commit removes about 150 LOC.
2016-02-16 17:47:50 +00:00
iter := &boundedIterator{
it: s.newIterator(pinnedChunkDescs, curriedQuarantineSeries, mss.evictRequests),
Streamline series iterator creation This will fix issue #1035 and will also help to make issue #1264 less bad. The fundamental problem in the current code: In the preload phase, we quite accurately determine which chunks will be used for the query being executed. However, in the subsequent step of creating series iterators, the created iterators are referencing _all_ in-memory chunks in their series, even the un-pinned ones. In iterator creation, we copy a pointer to each in-memory chunk of a series into the iterator. While this creates a certain amount of allocation churn, the worst thing about it is that copying the chunk pointer out of the chunkDesc requires a mutex acquisition. (Remember that the iterator will also reference un-pinned chunks, so we need to acquire the mutex to protect against concurrent eviction.) The worst case happens if a series doesn't even contain any relevant samples for the query time range. We notice that during preloading but then we will still create a series iterator for it. But even for series that do contain relevant samples, the overhead is quite bad for instant queries that retrieve a single sample from each series, but still go through all the effort of series iterator creation. All of that is particularly bad if a series has many in-memory chunks. This commit addresses the problem from two sides: First, it merges preloading and iterator creation into one step, i.e. the preload call returns an iterator for exactly the preloaded chunks. Second, the required mutex acquisition in chunkDesc has been greatly reduced. That was enabled by a side effect of the first step, which is that the iterator is only referencing pinned chunks, so there is no risk of concurrent eviction anymore, and chunks can be accessed without mutex acquisition. To simplify the code changes for the above, the long-planned change of ValueAtTime to ValueAtOrBefore time was performed at the same time. (It should have been done first, but it kind of accidentally happened while I was in the middle of writing the series iterator changes. Sorry for that.) So far, we actively filtered the up to two values that were returned by ValueAtTime, i.e. we invested work to retrieve up to two values, and then we invested more work to throw one of them away. The SeriesIterator.BoundaryValues method can be removed once #1401 is fixed. But I really didn't want to load even more changes into this PR. Benchmarks: The BenchmarkFuzz.* benchmarks run 83% faster (i.e. about six times faster) and allocate 95% fewer bytes. The reason for that is that the benchmark reads one sample after another from the time series and creates a new series iterator for each sample read. To find out how much these improvements matter in practice, I have mirrored a beefy Prometheus server at SoundCloud that suffers from both issues #1035 and #1264. To reach steady state that would be comparable, the server needs to run for 15d. So far, it has run for 1d. The test server currently has only half as many memory time series and 60% of the memory chunks the main server has. The 90th percentile rule evaluation cycle time is ~11s on the main server and only ~3s on the test server. However, these numbers might get much closer over time. In addition to performance improvements, this commit removes about 150 LOC.
2016-02-16 17:47:50 +00:00
start: model.Now().Add(-mss.dropAfter),
}
return iter, nil
Streamline series iterator creation This will fix issue #1035 and will also help to make issue #1264 less bad. The fundamental problem in the current code: In the preload phase, we quite accurately determine which chunks will be used for the query being executed. However, in the subsequent step of creating series iterators, the created iterators are referencing _all_ in-memory chunks in their series, even the un-pinned ones. In iterator creation, we copy a pointer to each in-memory chunk of a series into the iterator. While this creates a certain amount of allocation churn, the worst thing about it is that copying the chunk pointer out of the chunkDesc requires a mutex acquisition. (Remember that the iterator will also reference un-pinned chunks, so we need to acquire the mutex to protect against concurrent eviction.) The worst case happens if a series doesn't even contain any relevant samples for the query time range. We notice that during preloading but then we will still create a series iterator for it. But even for series that do contain relevant samples, the overhead is quite bad for instant queries that retrieve a single sample from each series, but still go through all the effort of series iterator creation. All of that is particularly bad if a series has many in-memory chunks. This commit addresses the problem from two sides: First, it merges preloading and iterator creation into one step, i.e. the preload call returns an iterator for exactly the preloaded chunks. Second, the required mutex acquisition in chunkDesc has been greatly reduced. That was enabled by a side effect of the first step, which is that the iterator is only referencing pinned chunks, so there is no risk of concurrent eviction anymore, and chunks can be accessed without mutex acquisition. To simplify the code changes for the above, the long-planned change of ValueAtTime to ValueAtOrBefore time was performed at the same time. (It should have been done first, but it kind of accidentally happened while I was in the middle of writing the series iterator changes. Sorry for that.) So far, we actively filtered the up to two values that were returned by ValueAtTime, i.e. we invested work to retrieve up to two values, and then we invested more work to throw one of them away. The SeriesIterator.BoundaryValues method can be removed once #1401 is fixed. But I really didn't want to load even more changes into this PR. Benchmarks: The BenchmarkFuzz.* benchmarks run 83% faster (i.e. about six times faster) and allocate 95% fewer bytes. The reason for that is that the benchmark reads one sample after another from the time series and creates a new series iterator for each sample read. To find out how much these improvements matter in practice, I have mirrored a beefy Prometheus server at SoundCloud that suffers from both issues #1035 and #1264. To reach steady state that would be comparable, the server needs to run for 15d. So far, it has run for 1d. The test server currently has only half as many memory time series and 60% of the memory chunks the main server has. The 90th percentile rule evaluation cycle time is ~11s on the main server and only ~3s on the test server. However, these numbers might get much closer over time. In addition to performance improvements, this commit removes about 150 LOC.
2016-02-16 17:47:50 +00:00
}
// newIterator returns a new SeriesIterator for the provided chunkDescs (which
// must be pinned).
//
// The caller must have locked the fingerprint of the memorySeries.
func (s *memorySeries) newIterator(
pinnedChunkDescs []*chunkDesc,
quarantine func(error),
evictRequests chan<- evictRequest,
) SeriesIterator {
Streamline series iterator creation This will fix issue #1035 and will also help to make issue #1264 less bad. The fundamental problem in the current code: In the preload phase, we quite accurately determine which chunks will be used for the query being executed. However, in the subsequent step of creating series iterators, the created iterators are referencing _all_ in-memory chunks in their series, even the un-pinned ones. In iterator creation, we copy a pointer to each in-memory chunk of a series into the iterator. While this creates a certain amount of allocation churn, the worst thing about it is that copying the chunk pointer out of the chunkDesc requires a mutex acquisition. (Remember that the iterator will also reference un-pinned chunks, so we need to acquire the mutex to protect against concurrent eviction.) The worst case happens if a series doesn't even contain any relevant samples for the query time range. We notice that during preloading but then we will still create a series iterator for it. But even for series that do contain relevant samples, the overhead is quite bad for instant queries that retrieve a single sample from each series, but still go through all the effort of series iterator creation. All of that is particularly bad if a series has many in-memory chunks. This commit addresses the problem from two sides: First, it merges preloading and iterator creation into one step, i.e. the preload call returns an iterator for exactly the preloaded chunks. Second, the required mutex acquisition in chunkDesc has been greatly reduced. That was enabled by a side effect of the first step, which is that the iterator is only referencing pinned chunks, so there is no risk of concurrent eviction anymore, and chunks can be accessed without mutex acquisition. To simplify the code changes for the above, the long-planned change of ValueAtTime to ValueAtOrBefore time was performed at the same time. (It should have been done first, but it kind of accidentally happened while I was in the middle of writing the series iterator changes. Sorry for that.) So far, we actively filtered the up to two values that were returned by ValueAtTime, i.e. we invested work to retrieve up to two values, and then we invested more work to throw one of them away. The SeriesIterator.BoundaryValues method can be removed once #1401 is fixed. But I really didn't want to load even more changes into this PR. Benchmarks: The BenchmarkFuzz.* benchmarks run 83% faster (i.e. about six times faster) and allocate 95% fewer bytes. The reason for that is that the benchmark reads one sample after another from the time series and creates a new series iterator for each sample read. To find out how much these improvements matter in practice, I have mirrored a beefy Prometheus server at SoundCloud that suffers from both issues #1035 and #1264. To reach steady state that would be comparable, the server needs to run for 15d. So far, it has run for 1d. The test server currently has only half as many memory time series and 60% of the memory chunks the main server has. The 90th percentile rule evaluation cycle time is ~11s on the main server and only ~3s on the test server. However, these numbers might get much closer over time. In addition to performance improvements, this commit removes about 150 LOC.
2016-02-16 17:47:50 +00:00
chunks := make([]chunk, 0, len(pinnedChunkDescs))
for _, cd := range pinnedChunkDescs {
// It's OK to directly access cd.c here (without locking) as the
// series FP is locked and the chunk is pinned.
chunks = append(chunks, cd.c)
}
return &memorySeriesIterator{
chunks: chunks,
chunkIts: make([]chunkIterator, len(chunks)),
quarantine: quarantine,
metric: s.metric,
pinnedChunkDescs: pinnedChunkDescs,
evictRequests: evictRequests,
Streamline series iterator creation This will fix issue #1035 and will also help to make issue #1264 less bad. The fundamental problem in the current code: In the preload phase, we quite accurately determine which chunks will be used for the query being executed. However, in the subsequent step of creating series iterators, the created iterators are referencing _all_ in-memory chunks in their series, even the un-pinned ones. In iterator creation, we copy a pointer to each in-memory chunk of a series into the iterator. While this creates a certain amount of allocation churn, the worst thing about it is that copying the chunk pointer out of the chunkDesc requires a mutex acquisition. (Remember that the iterator will also reference un-pinned chunks, so we need to acquire the mutex to protect against concurrent eviction.) The worst case happens if a series doesn't even contain any relevant samples for the query time range. We notice that during preloading but then we will still create a series iterator for it. But even for series that do contain relevant samples, the overhead is quite bad for instant queries that retrieve a single sample from each series, but still go through all the effort of series iterator creation. All of that is particularly bad if a series has many in-memory chunks. This commit addresses the problem from two sides: First, it merges preloading and iterator creation into one step, i.e. the preload call returns an iterator for exactly the preloaded chunks. Second, the required mutex acquisition in chunkDesc has been greatly reduced. That was enabled by a side effect of the first step, which is that the iterator is only referencing pinned chunks, so there is no risk of concurrent eviction anymore, and chunks can be accessed without mutex acquisition. To simplify the code changes for the above, the long-planned change of ValueAtTime to ValueAtOrBefore time was performed at the same time. (It should have been done first, but it kind of accidentally happened while I was in the middle of writing the series iterator changes. Sorry for that.) So far, we actively filtered the up to two values that were returned by ValueAtTime, i.e. we invested work to retrieve up to two values, and then we invested more work to throw one of them away. The SeriesIterator.BoundaryValues method can be removed once #1401 is fixed. But I really didn't want to load even more changes into this PR. Benchmarks: The BenchmarkFuzz.* benchmarks run 83% faster (i.e. about six times faster) and allocate 95% fewer bytes. The reason for that is that the benchmark reads one sample after another from the time series and creates a new series iterator for each sample read. To find out how much these improvements matter in practice, I have mirrored a beefy Prometheus server at SoundCloud that suffers from both issues #1035 and #1264. To reach steady state that would be comparable, the server needs to run for 15d. So far, it has run for 1d. The test server currently has only half as many memory time series and 60% of the memory chunks the main server has. The 90th percentile rule evaluation cycle time is ~11s on the main server and only ~3s on the test server. However, these numbers might get much closer over time. In addition to performance improvements, this commit removes about 150 LOC.
2016-02-16 17:47:50 +00:00
}
}
// preloadChunksForInstant preloads chunks for the latest value in the given
// range. If the last sample saved in the memorySeries itself is the latest
// value in the given range, it will in fact preload zero chunks and just take
// that value.
func (s *memorySeries) preloadChunksForInstant(
fp model.Fingerprint,
from model.Time, through model.Time,
mss *MemorySeriesStorage,
) (SeriesIterator, error) {
// If we have a lastSamplePair in the series, and thas last samplePair
// is in the interval, just take it in a singleSampleSeriesIterator. No
// need to pin or load anything.
lastSample := s.lastSamplePair()
if !through.Before(lastSample.Timestamp) &&
!from.After(lastSample.Timestamp) &&
lastSample != ZeroSamplePair {
iter := &boundedIterator{
it: &singleSampleSeriesIterator{
samplePair: lastSample,
metric: s.metric,
},
start: model.Now().Add(-mss.dropAfter),
}
return iter, nil
}
// If we are here, we are out of luck and have to delegate to the more
// expensive method.
return s.preloadChunksForRange(fp, from, through, mss)
}
// preloadChunksForRange loads chunks for the given range from the persistence.
// The caller must have locked the fingerprint of the series.
func (s *memorySeries) preloadChunksForRange(
fp model.Fingerprint,
from model.Time, through model.Time,
mss *MemorySeriesStorage,
) (SeriesIterator, error) {
firstChunkDescTime := model.Latest
if len(s.chunkDescs) > 0 {
firstChunkDescTime = s.chunkDescs[0].firstTime()
}
if s.chunkDescsOffset != 0 && from.Before(firstChunkDescTime) {
cds, err := mss.loadChunkDescs(fp, s.persistWatermark)
if err != nil {
return nopIter, err
}
s.chunkDescs = append(cds, s.chunkDescs...)
s.chunkDescsOffset = 0
s.persistWatermark += len(cds)
Streamline series iterator creation This will fix issue #1035 and will also help to make issue #1264 less bad. The fundamental problem in the current code: In the preload phase, we quite accurately determine which chunks will be used for the query being executed. However, in the subsequent step of creating series iterators, the created iterators are referencing _all_ in-memory chunks in their series, even the un-pinned ones. In iterator creation, we copy a pointer to each in-memory chunk of a series into the iterator. While this creates a certain amount of allocation churn, the worst thing about it is that copying the chunk pointer out of the chunkDesc requires a mutex acquisition. (Remember that the iterator will also reference un-pinned chunks, so we need to acquire the mutex to protect against concurrent eviction.) The worst case happens if a series doesn't even contain any relevant samples for the query time range. We notice that during preloading but then we will still create a series iterator for it. But even for series that do contain relevant samples, the overhead is quite bad for instant queries that retrieve a single sample from each series, but still go through all the effort of series iterator creation. All of that is particularly bad if a series has many in-memory chunks. This commit addresses the problem from two sides: First, it merges preloading and iterator creation into one step, i.e. the preload call returns an iterator for exactly the preloaded chunks. Second, the required mutex acquisition in chunkDesc has been greatly reduced. That was enabled by a side effect of the first step, which is that the iterator is only referencing pinned chunks, so there is no risk of concurrent eviction anymore, and chunks can be accessed without mutex acquisition. To simplify the code changes for the above, the long-planned change of ValueAtTime to ValueAtOrBefore time was performed at the same time. (It should have been done first, but it kind of accidentally happened while I was in the middle of writing the series iterator changes. Sorry for that.) So far, we actively filtered the up to two values that were returned by ValueAtTime, i.e. we invested work to retrieve up to two values, and then we invested more work to throw one of them away. The SeriesIterator.BoundaryValues method can be removed once #1401 is fixed. But I really didn't want to load even more changes into this PR. Benchmarks: The BenchmarkFuzz.* benchmarks run 83% faster (i.e. about six times faster) and allocate 95% fewer bytes. The reason for that is that the benchmark reads one sample after another from the time series and creates a new series iterator for each sample read. To find out how much these improvements matter in practice, I have mirrored a beefy Prometheus server at SoundCloud that suffers from both issues #1035 and #1264. To reach steady state that would be comparable, the server needs to run for 15d. So far, it has run for 1d. The test server currently has only half as many memory time series and 60% of the memory chunks the main server has. The 90th percentile rule evaluation cycle time is ~11s on the main server and only ~3s on the test server. However, these numbers might get much closer over time. In addition to performance improvements, this commit removes about 150 LOC.
2016-02-16 17:47:50 +00:00
firstChunkDescTime = s.chunkDescs[0].firstTime()
}
Streamline series iterator creation This will fix issue #1035 and will also help to make issue #1264 less bad. The fundamental problem in the current code: In the preload phase, we quite accurately determine which chunks will be used for the query being executed. However, in the subsequent step of creating series iterators, the created iterators are referencing _all_ in-memory chunks in their series, even the un-pinned ones. In iterator creation, we copy a pointer to each in-memory chunk of a series into the iterator. While this creates a certain amount of allocation churn, the worst thing about it is that copying the chunk pointer out of the chunkDesc requires a mutex acquisition. (Remember that the iterator will also reference un-pinned chunks, so we need to acquire the mutex to protect against concurrent eviction.) The worst case happens if a series doesn't even contain any relevant samples for the query time range. We notice that during preloading but then we will still create a series iterator for it. But even for series that do contain relevant samples, the overhead is quite bad for instant queries that retrieve a single sample from each series, but still go through all the effort of series iterator creation. All of that is particularly bad if a series has many in-memory chunks. This commit addresses the problem from two sides: First, it merges preloading and iterator creation into one step, i.e. the preload call returns an iterator for exactly the preloaded chunks. Second, the required mutex acquisition in chunkDesc has been greatly reduced. That was enabled by a side effect of the first step, which is that the iterator is only referencing pinned chunks, so there is no risk of concurrent eviction anymore, and chunks can be accessed without mutex acquisition. To simplify the code changes for the above, the long-planned change of ValueAtTime to ValueAtOrBefore time was performed at the same time. (It should have been done first, but it kind of accidentally happened while I was in the middle of writing the series iterator changes. Sorry for that.) So far, we actively filtered the up to two values that were returned by ValueAtTime, i.e. we invested work to retrieve up to two values, and then we invested more work to throw one of them away. The SeriesIterator.BoundaryValues method can be removed once #1401 is fixed. But I really didn't want to load even more changes into this PR. Benchmarks: The BenchmarkFuzz.* benchmarks run 83% faster (i.e. about six times faster) and allocate 95% fewer bytes. The reason for that is that the benchmark reads one sample after another from the time series and creates a new series iterator for each sample read. To find out how much these improvements matter in practice, I have mirrored a beefy Prometheus server at SoundCloud that suffers from both issues #1035 and #1264. To reach steady state that would be comparable, the server needs to run for 15d. So far, it has run for 1d. The test server currently has only half as many memory time series and 60% of the memory chunks the main server has. The 90th percentile rule evaluation cycle time is ~11s on the main server and only ~3s on the test server. However, these numbers might get much closer over time. In addition to performance improvements, this commit removes about 150 LOC.
2016-02-16 17:47:50 +00:00
if len(s.chunkDescs) == 0 || through.Before(firstChunkDescTime) {
return nopIter, nil
}
// Find first chunk with start time after "from".
fromIdx := sort.Search(len(s.chunkDescs), func(i int) bool {
return s.chunkDescs[i].firstTime().After(from)
})
// Find first chunk with start time after "through".
throughIdx := sort.Search(len(s.chunkDescs), func(i int) bool {
return s.chunkDescs[i].firstTime().After(through)
})
Streamline series iterator creation This will fix issue #1035 and will also help to make issue #1264 less bad. The fundamental problem in the current code: In the preload phase, we quite accurately determine which chunks will be used for the query being executed. However, in the subsequent step of creating series iterators, the created iterators are referencing _all_ in-memory chunks in their series, even the un-pinned ones. In iterator creation, we copy a pointer to each in-memory chunk of a series into the iterator. While this creates a certain amount of allocation churn, the worst thing about it is that copying the chunk pointer out of the chunkDesc requires a mutex acquisition. (Remember that the iterator will also reference un-pinned chunks, so we need to acquire the mutex to protect against concurrent eviction.) The worst case happens if a series doesn't even contain any relevant samples for the query time range. We notice that during preloading but then we will still create a series iterator for it. But even for series that do contain relevant samples, the overhead is quite bad for instant queries that retrieve a single sample from each series, but still go through all the effort of series iterator creation. All of that is particularly bad if a series has many in-memory chunks. This commit addresses the problem from two sides: First, it merges preloading and iterator creation into one step, i.e. the preload call returns an iterator for exactly the preloaded chunks. Second, the required mutex acquisition in chunkDesc has been greatly reduced. That was enabled by a side effect of the first step, which is that the iterator is only referencing pinned chunks, so there is no risk of concurrent eviction anymore, and chunks can be accessed without mutex acquisition. To simplify the code changes for the above, the long-planned change of ValueAtTime to ValueAtOrBefore time was performed at the same time. (It should have been done first, but it kind of accidentally happened while I was in the middle of writing the series iterator changes. Sorry for that.) So far, we actively filtered the up to two values that were returned by ValueAtTime, i.e. we invested work to retrieve up to two values, and then we invested more work to throw one of them away. The SeriesIterator.BoundaryValues method can be removed once #1401 is fixed. But I really didn't want to load even more changes into this PR. Benchmarks: The BenchmarkFuzz.* benchmarks run 83% faster (i.e. about six times faster) and allocate 95% fewer bytes. The reason for that is that the benchmark reads one sample after another from the time series and creates a new series iterator for each sample read. To find out how much these improvements matter in practice, I have mirrored a beefy Prometheus server at SoundCloud that suffers from both issues #1035 and #1264. To reach steady state that would be comparable, the server needs to run for 15d. So far, it has run for 1d. The test server currently has only half as many memory time series and 60% of the memory chunks the main server has. The 90th percentile rule evaluation cycle time is ~11s on the main server and only ~3s on the test server. However, these numbers might get much closer over time. In addition to performance improvements, this commit removes about 150 LOC.
2016-02-16 17:47:50 +00:00
if fromIdx == len(s.chunkDescs) {
// Even the last chunk starts before "from". Find out if the
// series ends before "from" and we don't need to do anything.
lt, err := s.chunkDescs[len(s.chunkDescs)-1].lastTime()
if err != nil {
return nopIter, err
}
if lt.Before(from) {
return nopIter, nil
Streamline series iterator creation This will fix issue #1035 and will also help to make issue #1264 less bad. The fundamental problem in the current code: In the preload phase, we quite accurately determine which chunks will be used for the query being executed. However, in the subsequent step of creating series iterators, the created iterators are referencing _all_ in-memory chunks in their series, even the un-pinned ones. In iterator creation, we copy a pointer to each in-memory chunk of a series into the iterator. While this creates a certain amount of allocation churn, the worst thing about it is that copying the chunk pointer out of the chunkDesc requires a mutex acquisition. (Remember that the iterator will also reference un-pinned chunks, so we need to acquire the mutex to protect against concurrent eviction.) The worst case happens if a series doesn't even contain any relevant samples for the query time range. We notice that during preloading but then we will still create a series iterator for it. But even for series that do contain relevant samples, the overhead is quite bad for instant queries that retrieve a single sample from each series, but still go through all the effort of series iterator creation. All of that is particularly bad if a series has many in-memory chunks. This commit addresses the problem from two sides: First, it merges preloading and iterator creation into one step, i.e. the preload call returns an iterator for exactly the preloaded chunks. Second, the required mutex acquisition in chunkDesc has been greatly reduced. That was enabled by a side effect of the first step, which is that the iterator is only referencing pinned chunks, so there is no risk of concurrent eviction anymore, and chunks can be accessed without mutex acquisition. To simplify the code changes for the above, the long-planned change of ValueAtTime to ValueAtOrBefore time was performed at the same time. (It should have been done first, but it kind of accidentally happened while I was in the middle of writing the series iterator changes. Sorry for that.) So far, we actively filtered the up to two values that were returned by ValueAtTime, i.e. we invested work to retrieve up to two values, and then we invested more work to throw one of them away. The SeriesIterator.BoundaryValues method can be removed once #1401 is fixed. But I really didn't want to load even more changes into this PR. Benchmarks: The BenchmarkFuzz.* benchmarks run 83% faster (i.e. about six times faster) and allocate 95% fewer bytes. The reason for that is that the benchmark reads one sample after another from the time series and creates a new series iterator for each sample read. To find out how much these improvements matter in practice, I have mirrored a beefy Prometheus server at SoundCloud that suffers from both issues #1035 and #1264. To reach steady state that would be comparable, the server needs to run for 15d. So far, it has run for 1d. The test server currently has only half as many memory time series and 60% of the memory chunks the main server has. The 90th percentile rule evaluation cycle time is ~11s on the main server and only ~3s on the test server. However, these numbers might get much closer over time. In addition to performance improvements, this commit removes about 150 LOC.
2016-02-16 17:47:50 +00:00
}
}
if fromIdx > 0 {
fromIdx--
}
if throughIdx == len(s.chunkDescs) {
throughIdx--
}
pinIndexes := make([]int, 0, throughIdx-fromIdx+1)
for i := fromIdx; i <= throughIdx; i++ {
pinIndexes = append(pinIndexes, i)
}
2015-05-06 14:53:12 +00:00
return s.preloadChunks(pinIndexes, fp, mss)
}
// head returns a pointer to the head chunk descriptor. The caller must have
// locked the fingerprint of the memorySeries. This method will panic if this
// series has no chunk descriptors.
func (s *memorySeries) head() *chunkDesc {
return s.chunkDescs[len(s.chunkDescs)-1]
}
// firstTime returns the timestamp of the first sample in the series.
//
// The caller must have locked the fingerprint of the memorySeries.
func (s *memorySeries) firstTime() model.Time {
if s.chunkDescsOffset == 0 && len(s.chunkDescs) > 0 {
return s.chunkDescs[0].firstTime()
}
return s.savedFirstTime
}
// lastSamplePair returns the last ingested SamplePair. It returns
// ZeroSamplePair if this memorySeries has never received a sample (via the add
// method), which is the case for freshly unarchived series or newly created
// ones and also for all series after a server restart. However, in that case,
// series will most likely be considered stale anyway.
//
// The caller must have locked the fingerprint of the memorySeries.
func (s *memorySeries) lastSamplePair() model.SamplePair {
if !s.lastSampleValueSet {
return ZeroSamplePair
}
return model.SamplePair{
Timestamp: s.lastTime,
Value: s.lastSampleValue,
}
}
// chunksToPersist returns a slice of chunkDescs eligible for persistence. It's
// the caller's responsibility to actually persist the returned chunks
// afterwards. The method sets the persistWatermark and the dirty flag
// accordingly.
//
// The caller must have locked the fingerprint of the series.
func (s *memorySeries) chunksToPersist() []*chunkDesc {
newWatermark := len(s.chunkDescs)
if !s.headChunkClosed {
newWatermark--
}
if newWatermark == s.persistWatermark {
return nil
}
cds := s.chunkDescs[s.persistWatermark:newWatermark]
s.dirty = true
s.persistWatermark = newWatermark
return cds
}
// memorySeriesIterator implements SeriesIterator.
type memorySeriesIterator struct {
// Last chunkIterator used by ValueAtOrBeforeTime.
chunkIt chunkIterator
// Caches chunkIterators.
chunkIts []chunkIterator
// The actual sample chunks.
chunks []chunk
// Call to quarantine the series this iterator belongs to.
quarantine func(error)
// The metric corresponding to the iterator.
metric model.Metric
// Chunks that were pinned for this iterator.
pinnedChunkDescs []*chunkDesc
// Where to send evict requests when unpinning pinned chunks.
evictRequests chan<- evictRequest
}
Streamline series iterator creation This will fix issue #1035 and will also help to make issue #1264 less bad. The fundamental problem in the current code: In the preload phase, we quite accurately determine which chunks will be used for the query being executed. However, in the subsequent step of creating series iterators, the created iterators are referencing _all_ in-memory chunks in their series, even the un-pinned ones. In iterator creation, we copy a pointer to each in-memory chunk of a series into the iterator. While this creates a certain amount of allocation churn, the worst thing about it is that copying the chunk pointer out of the chunkDesc requires a mutex acquisition. (Remember that the iterator will also reference un-pinned chunks, so we need to acquire the mutex to protect against concurrent eviction.) The worst case happens if a series doesn't even contain any relevant samples for the query time range. We notice that during preloading but then we will still create a series iterator for it. But even for series that do contain relevant samples, the overhead is quite bad for instant queries that retrieve a single sample from each series, but still go through all the effort of series iterator creation. All of that is particularly bad if a series has many in-memory chunks. This commit addresses the problem from two sides: First, it merges preloading and iterator creation into one step, i.e. the preload call returns an iterator for exactly the preloaded chunks. Second, the required mutex acquisition in chunkDesc has been greatly reduced. That was enabled by a side effect of the first step, which is that the iterator is only referencing pinned chunks, so there is no risk of concurrent eviction anymore, and chunks can be accessed without mutex acquisition. To simplify the code changes for the above, the long-planned change of ValueAtTime to ValueAtOrBefore time was performed at the same time. (It should have been done first, but it kind of accidentally happened while I was in the middle of writing the series iterator changes. Sorry for that.) So far, we actively filtered the up to two values that were returned by ValueAtTime, i.e. we invested work to retrieve up to two values, and then we invested more work to throw one of them away. The SeriesIterator.BoundaryValues method can be removed once #1401 is fixed. But I really didn't want to load even more changes into this PR. Benchmarks: The BenchmarkFuzz.* benchmarks run 83% faster (i.e. about six times faster) and allocate 95% fewer bytes. The reason for that is that the benchmark reads one sample after another from the time series and creates a new series iterator for each sample read. To find out how much these improvements matter in practice, I have mirrored a beefy Prometheus server at SoundCloud that suffers from both issues #1035 and #1264. To reach steady state that would be comparable, the server needs to run for 15d. So far, it has run for 1d. The test server currently has only half as many memory time series and 60% of the memory chunks the main server has. The 90th percentile rule evaluation cycle time is ~11s on the main server and only ~3s on the test server. However, these numbers might get much closer over time. In addition to performance improvements, this commit removes about 150 LOC.
2016-02-16 17:47:50 +00:00
// ValueAtOrBeforeTime implements SeriesIterator.
func (it *memorySeriesIterator) ValueAtOrBeforeTime(t model.Time) model.SamplePair {
// The most common case. We are iterating through a chunk.
if it.chunkIt != nil {
containsT, err := it.chunkIt.contains(t)
if err != nil {
it.quarantine(err)
return ZeroSamplePair
}
if containsT {
2016-03-09 15:20:39 +00:00
if it.chunkIt.findAtOrBefore(t) {
return it.chunkIt.value()
}
2016-03-09 15:20:39 +00:00
if it.chunkIt.err() != nil {
it.quarantine(it.chunkIt.err())
}
return ZeroSamplePair
}
}
if len(it.chunks) == 0 {
return ZeroSamplePair
}
Streamline series iterator creation This will fix issue #1035 and will also help to make issue #1264 less bad. The fundamental problem in the current code: In the preload phase, we quite accurately determine which chunks will be used for the query being executed. However, in the subsequent step of creating series iterators, the created iterators are referencing _all_ in-memory chunks in their series, even the un-pinned ones. In iterator creation, we copy a pointer to each in-memory chunk of a series into the iterator. While this creates a certain amount of allocation churn, the worst thing about it is that copying the chunk pointer out of the chunkDesc requires a mutex acquisition. (Remember that the iterator will also reference un-pinned chunks, so we need to acquire the mutex to protect against concurrent eviction.) The worst case happens if a series doesn't even contain any relevant samples for the query time range. We notice that during preloading but then we will still create a series iterator for it. But even for series that do contain relevant samples, the overhead is quite bad for instant queries that retrieve a single sample from each series, but still go through all the effort of series iterator creation. All of that is particularly bad if a series has many in-memory chunks. This commit addresses the problem from two sides: First, it merges preloading and iterator creation into one step, i.e. the preload call returns an iterator for exactly the preloaded chunks. Second, the required mutex acquisition in chunkDesc has been greatly reduced. That was enabled by a side effect of the first step, which is that the iterator is only referencing pinned chunks, so there is no risk of concurrent eviction anymore, and chunks can be accessed without mutex acquisition. To simplify the code changes for the above, the long-planned change of ValueAtTime to ValueAtOrBefore time was performed at the same time. (It should have been done first, but it kind of accidentally happened while I was in the middle of writing the series iterator changes. Sorry for that.) So far, we actively filtered the up to two values that were returned by ValueAtTime, i.e. we invested work to retrieve up to two values, and then we invested more work to throw one of them away. The SeriesIterator.BoundaryValues method can be removed once #1401 is fixed. But I really didn't want to load even more changes into this PR. Benchmarks: The BenchmarkFuzz.* benchmarks run 83% faster (i.e. about six times faster) and allocate 95% fewer bytes. The reason for that is that the benchmark reads one sample after another from the time series and creates a new series iterator for each sample read. To find out how much these improvements matter in practice, I have mirrored a beefy Prometheus server at SoundCloud that suffers from both issues #1035 and #1264. To reach steady state that would be comparable, the server needs to run for 15d. So far, it has run for 1d. The test server currently has only half as many memory time series and 60% of the memory chunks the main server has. The 90th percentile rule evaluation cycle time is ~11s on the main server and only ~3s on the test server. However, these numbers might get much closer over time. In addition to performance improvements, this commit removes about 150 LOC.
2016-02-16 17:47:50 +00:00
// Find the last chunk where firstTime() is before or equal to t.
l := len(it.chunks) - 1
i := sort.Search(len(it.chunks), func(i int) bool {
return !it.chunks[l-i].firstTime().After(t)
})
if i == len(it.chunks) {
Streamline series iterator creation This will fix issue #1035 and will also help to make issue #1264 less bad. The fundamental problem in the current code: In the preload phase, we quite accurately determine which chunks will be used for the query being executed. However, in the subsequent step of creating series iterators, the created iterators are referencing _all_ in-memory chunks in their series, even the un-pinned ones. In iterator creation, we copy a pointer to each in-memory chunk of a series into the iterator. While this creates a certain amount of allocation churn, the worst thing about it is that copying the chunk pointer out of the chunkDesc requires a mutex acquisition. (Remember that the iterator will also reference un-pinned chunks, so we need to acquire the mutex to protect against concurrent eviction.) The worst case happens if a series doesn't even contain any relevant samples for the query time range. We notice that during preloading but then we will still create a series iterator for it. But even for series that do contain relevant samples, the overhead is quite bad for instant queries that retrieve a single sample from each series, but still go through all the effort of series iterator creation. All of that is particularly bad if a series has many in-memory chunks. This commit addresses the problem from two sides: First, it merges preloading and iterator creation into one step, i.e. the preload call returns an iterator for exactly the preloaded chunks. Second, the required mutex acquisition in chunkDesc has been greatly reduced. That was enabled by a side effect of the first step, which is that the iterator is only referencing pinned chunks, so there is no risk of concurrent eviction anymore, and chunks can be accessed without mutex acquisition. To simplify the code changes for the above, the long-planned change of ValueAtTime to ValueAtOrBefore time was performed at the same time. (It should have been done first, but it kind of accidentally happened while I was in the middle of writing the series iterator changes. Sorry for that.) So far, we actively filtered the up to two values that were returned by ValueAtTime, i.e. we invested work to retrieve up to two values, and then we invested more work to throw one of them away. The SeriesIterator.BoundaryValues method can be removed once #1401 is fixed. But I really didn't want to load even more changes into this PR. Benchmarks: The BenchmarkFuzz.* benchmarks run 83% faster (i.e. about six times faster) and allocate 95% fewer bytes. The reason for that is that the benchmark reads one sample after another from the time series and creates a new series iterator for each sample read. To find out how much these improvements matter in practice, I have mirrored a beefy Prometheus server at SoundCloud that suffers from both issues #1035 and #1264. To reach steady state that would be comparable, the server needs to run for 15d. So far, it has run for 1d. The test server currently has only half as many memory time series and 60% of the memory chunks the main server has. The 90th percentile rule evaluation cycle time is ~11s on the main server and only ~3s on the test server. However, these numbers might get much closer over time. In addition to performance improvements, this commit removes about 150 LOC.
2016-02-16 17:47:50 +00:00
// Even the first chunk starts after t.
return ZeroSamplePair
}
it.chunkIt = it.chunkIterator(l - i)
2016-03-09 15:20:39 +00:00
if it.chunkIt.findAtOrBefore(t) {
return it.chunkIt.value()
}
if it.chunkIt.err() != nil {
it.quarantine(it.chunkIt.err())
}
2016-03-09 15:20:39 +00:00
return ZeroSamplePair
}
// RangeValues implements SeriesIterator.
func (it *memorySeriesIterator) RangeValues(in metric.Interval) []model.SamplePair {
// Find the first chunk for which the first sample is within the interval.
i := sort.Search(len(it.chunks), func(i int) bool {
return !it.chunks[i].firstTime().Before(in.OldestInclusive)
})
// Only now check the last timestamp of the previous chunk (which is
// fairly expensive).
if i > 0 {
lt, err := it.chunkIterator(i - 1).lastTimestamp()
if err != nil {
it.quarantine(err)
return nil
}
if !lt.Before(in.OldestInclusive) {
i--
}
}
values := []model.SamplePair{}
for j, c := range it.chunks[i:] {
if c.firstTime().After(in.NewestInclusive) {
break
}
2016-03-09 15:20:39 +00:00
chValues, err := rangeValues(it.chunkIterator(i+j), in)
if err != nil {
it.quarantine(err)
return nil
}
values = append(values, chValues...)
}
return values
}
func (it *memorySeriesIterator) Metric() metric.Metric {
return metric.Metric{Metric: it.metric}
}
// chunkIterator returns the chunkIterator for the chunk at position i (and
// creates it if needed).
func (it *memorySeriesIterator) chunkIterator(i int) chunkIterator {
chunkIt := it.chunkIts[i]
if chunkIt == nil {
chunkIt = it.chunks[i].newIterator()
it.chunkIts[i] = chunkIt
}
return chunkIt
}
func (it *memorySeriesIterator) Close() {
for _, cd := range it.pinnedChunkDescs {
cd.unpin(it.evictRequests)
}
chunkOps.WithLabelValues(unpin).Add(float64(len(it.pinnedChunkDescs)))
}
// singleSampleSeriesIterator implements Series Iterator. It is a "shortcut
// iterator" that returns a single sample only. The sample is saved in the
// iterator itself, so no chunks need to be pinned.
type singleSampleSeriesIterator struct {
samplePair model.SamplePair
metric model.Metric
}
// ValueAtTime implements SeriesIterator.
func (it *singleSampleSeriesIterator) ValueAtOrBeforeTime(t model.Time) model.SamplePair {
if it.samplePair.Timestamp.After(t) {
return ZeroSamplePair
}
return it.samplePair
}
// RangeValues implements SeriesIterator.
func (it *singleSampleSeriesIterator) RangeValues(in metric.Interval) []model.SamplePair {
if it.samplePair.Timestamp.After(in.NewestInclusive) ||
it.samplePair.Timestamp.Before(in.OldestInclusive) {
return []model.SamplePair{}
}
return []model.SamplePair{it.samplePair}
}
func (it *singleSampleSeriesIterator) Metric() metric.Metric {
return metric.Metric{Metric: it.metric}
}
// Close implements SeriesIterator.
func (it *singleSampleSeriesIterator) Close() {}
// nopSeriesIterator implements Series Iterator. It never returns any values.
type nopSeriesIterator struct{}
// ValueAtTime implements SeriesIterator.
Streamline series iterator creation This will fix issue #1035 and will also help to make issue #1264 less bad. The fundamental problem in the current code: In the preload phase, we quite accurately determine which chunks will be used for the query being executed. However, in the subsequent step of creating series iterators, the created iterators are referencing _all_ in-memory chunks in their series, even the un-pinned ones. In iterator creation, we copy a pointer to each in-memory chunk of a series into the iterator. While this creates a certain amount of allocation churn, the worst thing about it is that copying the chunk pointer out of the chunkDesc requires a mutex acquisition. (Remember that the iterator will also reference un-pinned chunks, so we need to acquire the mutex to protect against concurrent eviction.) The worst case happens if a series doesn't even contain any relevant samples for the query time range. We notice that during preloading but then we will still create a series iterator for it. But even for series that do contain relevant samples, the overhead is quite bad for instant queries that retrieve a single sample from each series, but still go through all the effort of series iterator creation. All of that is particularly bad if a series has many in-memory chunks. This commit addresses the problem from two sides: First, it merges preloading and iterator creation into one step, i.e. the preload call returns an iterator for exactly the preloaded chunks. Second, the required mutex acquisition in chunkDesc has been greatly reduced. That was enabled by a side effect of the first step, which is that the iterator is only referencing pinned chunks, so there is no risk of concurrent eviction anymore, and chunks can be accessed without mutex acquisition. To simplify the code changes for the above, the long-planned change of ValueAtTime to ValueAtOrBefore time was performed at the same time. (It should have been done first, but it kind of accidentally happened while I was in the middle of writing the series iterator changes. Sorry for that.) So far, we actively filtered the up to two values that were returned by ValueAtTime, i.e. we invested work to retrieve up to two values, and then we invested more work to throw one of them away. The SeriesIterator.BoundaryValues method can be removed once #1401 is fixed. But I really didn't want to load even more changes into this PR. Benchmarks: The BenchmarkFuzz.* benchmarks run 83% faster (i.e. about six times faster) and allocate 95% fewer bytes. The reason for that is that the benchmark reads one sample after another from the time series and creates a new series iterator for each sample read. To find out how much these improvements matter in practice, I have mirrored a beefy Prometheus server at SoundCloud that suffers from both issues #1035 and #1264. To reach steady state that would be comparable, the server needs to run for 15d. So far, it has run for 1d. The test server currently has only half as many memory time series and 60% of the memory chunks the main server has. The 90th percentile rule evaluation cycle time is ~11s on the main server and only ~3s on the test server. However, these numbers might get much closer over time. In addition to performance improvements, this commit removes about 150 LOC.
2016-02-16 17:47:50 +00:00
func (i nopSeriesIterator) ValueAtOrBeforeTime(t model.Time) model.SamplePair {
return ZeroSamplePair
}
// RangeValues implements SeriesIterator.
func (i nopSeriesIterator) RangeValues(in metric.Interval) []model.SamplePair {
return []model.SamplePair{}
}
Streamline series iterator creation This will fix issue #1035 and will also help to make issue #1264 less bad. The fundamental problem in the current code: In the preload phase, we quite accurately determine which chunks will be used for the query being executed. However, in the subsequent step of creating series iterators, the created iterators are referencing _all_ in-memory chunks in their series, even the un-pinned ones. In iterator creation, we copy a pointer to each in-memory chunk of a series into the iterator. While this creates a certain amount of allocation churn, the worst thing about it is that copying the chunk pointer out of the chunkDesc requires a mutex acquisition. (Remember that the iterator will also reference un-pinned chunks, so we need to acquire the mutex to protect against concurrent eviction.) The worst case happens if a series doesn't even contain any relevant samples for the query time range. We notice that during preloading but then we will still create a series iterator for it. But even for series that do contain relevant samples, the overhead is quite bad for instant queries that retrieve a single sample from each series, but still go through all the effort of series iterator creation. All of that is particularly bad if a series has many in-memory chunks. This commit addresses the problem from two sides: First, it merges preloading and iterator creation into one step, i.e. the preload call returns an iterator for exactly the preloaded chunks. Second, the required mutex acquisition in chunkDesc has been greatly reduced. That was enabled by a side effect of the first step, which is that the iterator is only referencing pinned chunks, so there is no risk of concurrent eviction anymore, and chunks can be accessed without mutex acquisition. To simplify the code changes for the above, the long-planned change of ValueAtTime to ValueAtOrBefore time was performed at the same time. (It should have been done first, but it kind of accidentally happened while I was in the middle of writing the series iterator changes. Sorry for that.) So far, we actively filtered the up to two values that were returned by ValueAtTime, i.e. we invested work to retrieve up to two values, and then we invested more work to throw one of them away. The SeriesIterator.BoundaryValues method can be removed once #1401 is fixed. But I really didn't want to load even more changes into this PR. Benchmarks: The BenchmarkFuzz.* benchmarks run 83% faster (i.e. about six times faster) and allocate 95% fewer bytes. The reason for that is that the benchmark reads one sample after another from the time series and creates a new series iterator for each sample read. To find out how much these improvements matter in practice, I have mirrored a beefy Prometheus server at SoundCloud that suffers from both issues #1035 and #1264. To reach steady state that would be comparable, the server needs to run for 15d. So far, it has run for 1d. The test server currently has only half as many memory time series and 60% of the memory chunks the main server has. The 90th percentile rule evaluation cycle time is ~11s on the main server and only ~3s on the test server. However, these numbers might get much closer over time. In addition to performance improvements, this commit removes about 150 LOC.
2016-02-16 17:47:50 +00:00
// Metric implements SeriesIterator.
func (i nopSeriesIterator) Metric() metric.Metric {
return metric.Metric{}
}
// Close implements SeriesIterator.
func (i nopSeriesIterator) Close() {}
Streamline series iterator creation This will fix issue #1035 and will also help to make issue #1264 less bad. The fundamental problem in the current code: In the preload phase, we quite accurately determine which chunks will be used for the query being executed. However, in the subsequent step of creating series iterators, the created iterators are referencing _all_ in-memory chunks in their series, even the un-pinned ones. In iterator creation, we copy a pointer to each in-memory chunk of a series into the iterator. While this creates a certain amount of allocation churn, the worst thing about it is that copying the chunk pointer out of the chunkDesc requires a mutex acquisition. (Remember that the iterator will also reference un-pinned chunks, so we need to acquire the mutex to protect against concurrent eviction.) The worst case happens if a series doesn't even contain any relevant samples for the query time range. We notice that during preloading but then we will still create a series iterator for it. But even for series that do contain relevant samples, the overhead is quite bad for instant queries that retrieve a single sample from each series, but still go through all the effort of series iterator creation. All of that is particularly bad if a series has many in-memory chunks. This commit addresses the problem from two sides: First, it merges preloading and iterator creation into one step, i.e. the preload call returns an iterator for exactly the preloaded chunks. Second, the required mutex acquisition in chunkDesc has been greatly reduced. That was enabled by a side effect of the first step, which is that the iterator is only referencing pinned chunks, so there is no risk of concurrent eviction anymore, and chunks can be accessed without mutex acquisition. To simplify the code changes for the above, the long-planned change of ValueAtTime to ValueAtOrBefore time was performed at the same time. (It should have been done first, but it kind of accidentally happened while I was in the middle of writing the series iterator changes. Sorry for that.) So far, we actively filtered the up to two values that were returned by ValueAtTime, i.e. we invested work to retrieve up to two values, and then we invested more work to throw one of them away. The SeriesIterator.BoundaryValues method can be removed once #1401 is fixed. But I really didn't want to load even more changes into this PR. Benchmarks: The BenchmarkFuzz.* benchmarks run 83% faster (i.e. about six times faster) and allocate 95% fewer bytes. The reason for that is that the benchmark reads one sample after another from the time series and creates a new series iterator for each sample read. To find out how much these improvements matter in practice, I have mirrored a beefy Prometheus server at SoundCloud that suffers from both issues #1035 and #1264. To reach steady state that would be comparable, the server needs to run for 15d. So far, it has run for 1d. The test server currently has only half as many memory time series and 60% of the memory chunks the main server has. The 90th percentile rule evaluation cycle time is ~11s on the main server and only ~3s on the test server. However, these numbers might get much closer over time. In addition to performance improvements, this commit removes about 150 LOC.
2016-02-16 17:47:50 +00:00
var nopIter nopSeriesIterator // A nopSeriesIterator for convenience. Can be shared.