Merge pull request #1345 from prometheus/beorn7/storage2

Fix multiple issues around chunks to persist.

cmd/prometheus/config.go

@@ -113,7 +113,7 @@ func init() {
 		"How long to retain samples in the local storage.",
 	)
 	cfg.fs.IntVar(
-		&cfg.storage.MaxChunksToPersist, "storage.local.max-chunks-to-persist", 1024*1024,
+		&cfg.storage.MaxChunksToPersist, "storage.local.max-chunks-to-persist", 512*1024,
 		"How many chunks can be waiting for persistence before sample ingestion will stop. Many chunks waiting to be persisted will increase the checkpoint size.",
 	)
 	cfg.fs.DurationVar(

storage/local/persistence.go

@@ -861,6 +861,12 @@ func (p *persistence) loadSeriesMapAndHeads() (sm *seriesMap, chunksToPersist in
 			}
 		}
 
+		headChunkClosed := persistWatermark >= numChunkDescs
+		if !headChunkClosed {
+			// Head chunk is not ready for persisting yet.
+			chunksToPersist--
+		}
+
 		fingerprintToSeries[model.Fingerprint(fp)] = &memorySeries{
 			metric:           model.Metric(metric),
 			chunkDescs:       chunkDescs,
@@ -869,7 +875,7 @@ func (p *persistence) loadSeriesMapAndHeads() (sm *seriesMap, chunksToPersist in
 			chunkDescsOffset: int(chunkDescsOffset),
 			savedFirstTime:   model.Time(savedFirstTime),
 			lastTime:         chunkDescs[len(chunkDescs)-1].lastTime(),
-			headChunkClosed:  persistWatermark >= numChunkDescs,
+			headChunkClosed:  headChunkClosed,
 		}
 	}
 	return sm, chunksToPersist, nil

storage/local/storage.go

@@ -17,6 +17,7 @@ package local
 import (
 	"container/list"
 	"fmt"
+	"math"
 	"sync/atomic"
 	"time"
 
@@ -38,11 +39,23 @@ const (
 	// See waitForNextFP.
 	maxEvictInterval = time.Minute
 
-	// If numChunskToPersist is this percentage of maxChunksToPersist, we
-	// consider the storage in "graceful degradation mode", i.e. we do not
-	// checkpoint anymore based on the dirty series count, and we do not
-	// sync series files anymore if using the adaptive sync strategy.
-	percentChunksToPersistForDegradation = 80
+	// Constants to control the hysteresis of entering and leaving "rushed
+	// mode". In rushed mode, the dirty series count is ignored for
+	// checkpointing, series are maintained as frequently as possible, and
+	// series files are not synced if the adaptive sync strategy is used.
+	persintenceUrgencyScoreForEnteringRushedMode = 0.8
+	persintenceUrgencyScoreForLeavingRushedMode  = 0.7
+
+	// This factor times -storage.local.memory-chunks is the number of
+	// memory chunks we tolerate before suspending ingestion (TODO!). It is
+	// also a basis for calculating the persistenceUrgencyScore.
+	toleranceFactorForMemChunks = 1.1
+	// This factor times -storage.local.max-chunks-to-persist is the minimum
+	// required number of chunks waiting for persistence before the number
+	// of chunks in memory may influence the persistenceUrgencyScore. (In
+	// other words: if there are no chunks to persist, it doesn't help chunk
+	// eviction if we speed up persistence.)
+	factorMinChunksToPersist = 0.2
 )
 
 var (
@@ -110,7 +123,7 @@ type memorySeriesStorage struct {
 	// numChunksToPersist has to be aligned for atomic operations.
 	numChunksToPersist int64 // The number of chunks waiting for persistence.
 	maxChunksToPersist int   // If numChunksToPersist reaches this threshold, ingestion will stall.
-	degraded           bool
+	rushed             bool  // Whether the storage is in rushed mode.
 
 	fpLocker   *fingerprintLocker
 	fpToSeries *seriesMap
@@ -137,6 +150,8 @@ type memorySeriesStorage struct {
 	outOfOrderSamplesCount      prometheus.Counter
 	invalidPreloadRequestsCount prometheus.Counter
 	maintainSeriesDuration      *prometheus.SummaryVec
+	persistenceUrgencyScore     prometheus.Gauge
+	rushedMode                  prometheus.Gauge
 }
 
 // MemorySeriesStorageOptions contains options needed by
@@ -225,6 +240,18 @@ func NewMemorySeriesStorage(o *MemorySeriesStorageOptions) Storage {
 			},
 			[]string{seriesLocationLabel},
 		),
+		persistenceUrgencyScore: prometheus.NewGauge(prometheus.GaugeOpts{
+			Namespace: namespace,
+			Subsystem: subsystem,
+			Name:      "persistence_urgency_score",
+			Help:      "A score of urgency to persist chunks, 0 is least urgent, 1 most.",
+		}),
+		rushedMode: prometheus.NewGauge(prometheus.GaugeOpts{
+			Namespace: namespace,
+			Subsystem: subsystem,
+			Name:      "rushed_mode",
+			Help:      "1 if the storage is in rushed mode, 0 otherwise. In rushed mode, the system behaves as if the persistence_urgency_score is 1.",
+		}),
 	}
 	return s
 }
@@ -238,7 +265,7 @@ func (s *memorySeriesStorage) Start() (err error) {
 	case Always:
 		syncStrategy = func() bool { return true }
 	case Adaptive:
-		syncStrategy = func() bool { return !s.isDegraded() }
+		syncStrategy = func() bool { return s.calculatePersistenceUrgencyScore() < 1 }
 	default:
 		panic("unknown sync strategy")
 	}
@@ -805,8 +832,8 @@ func (s *memorySeriesStorage) cycleThroughMemoryFingerprints() chan model.Finger
 				case <-s.loopStopping:
 					return
 				}
-				// Reduce the wait time by the backlog score.
-				s.waitForNextFP(s.fpToSeries.length(), s.persistenceBacklogScore())
+				// Reduce the wait time according to the urgency score.
+				s.waitForNextFP(s.fpToSeries.length(), 1-s.calculatePersistenceUrgencyScore())
 				count++
 			}
 			if count > 0 {
@@ -898,9 +925,9 @@ loop:
 				// would be counterproductive, as it would slow down chunk persisting even more,
 				// while in a situation like that, where we are clearly lacking speed of disk
 				// maintenance, the best we can do for crash recovery is to persist chunks as
-				// quickly as possible. So only checkpoint if the storage is not in "graceful
-				// degradation mode".
-				if dirtySeriesCount >= s.checkpointDirtySeriesLimit && !s.isDegraded() {
+				// quickly as possible. So only checkpoint if the urgency score is < 1.
+				if dirtySeriesCount >= s.checkpointDirtySeriesLimit &&
+					s.calculatePersistenceUrgencyScore() < 1 {
 					checkpointTimer.Reset(0)
 				}
 			}
@@ -1144,36 +1171,84 @@ func (s *memorySeriesStorage) incNumChunksToPersist(by int) {
 	atomic.AddInt64(&s.numChunksToPersist, int64(by))
 }
 
-// isDegraded returns whether the storage is in "graceful degradation mode",
-// which is the case if the number of chunks waiting for persistence has reached
-// a percentage of maxChunksToPersist that exceeds
-// percentChunksToPersistForDegradation. The method is not goroutine safe (but
-// only ever called from the goroutine dealing with series maintenance).
-// Changes of degradation mode are logged.
-func (s *memorySeriesStorage) isDegraded() bool {
-	nowDegraded := s.getNumChunksToPersist() > s.maxChunksToPersist*percentChunksToPersistForDegradation/100
-	if s.degraded && !nowDegraded {
-		log.Warn("Storage has left graceful degradation mode. Things are back to normal.")
-	} else if !s.degraded && nowDegraded {
-		log.Warnf(
-			"%d chunks waiting for persistence (%d%% of the allowed maximum %d). Storage is now in graceful degradation mode. Series files are not synced anymore if following the adaptive strategy. Checkpoints are not performed more often than every %v. Series maintenance happens as frequently as possible.",
-			s.getNumChunksToPersist(),
-			s.getNumChunksToPersist()*100/s.maxChunksToPersist,
-			s.maxChunksToPersist,
-			s.checkpointInterval)
+// calculatePersistenceUrgencyScore calculates and returns an urgency score for
+// the speed of persisting chunks. The score is between 0 and 1, where 0 means
+// no urgency at all and 1 means highest urgency.
+//
+// The score is the maximum of the two following sub-scores:
+//
+// (1) The first sub-score is the number of chunks waiting for persistence
+// divided by the maximum number of chunks allowed to be waiting for
+// persistence.
+//
+// (2) If there are more chunks in memory than allowed AND there are more chunks
+// waiting for persistence than factorMinChunksToPersist times
+// -storage.local.max-chunks-to-persist, then the second sub-score is the
+// fraction the number of memory chunks has reached between
+// -storage.local.memory-chunks and toleranceFactorForMemChunks times
+// -storage.local.memory-chunks.
+//
+// Should the score ever hit persintenceUrgencyScoreForEnteringRushedMode, the
+// storage locks into "rushed mode", in which the returned score is always
+// bumped up to 1 until the non-bumped score is below
+// persintenceUrgencyScoreForLeavingRushedMode.
+//
+// This method is not goroutine-safe, but it is only ever called by the single
+// goroutine that is in charge of series maintenance. According to the returned
+// score, series maintenence should be sped up. If a score of 1 is returned,
+// checkpointing based on dirty-series count should be disabled, and series
+// files should not by synced anymore provided the user has specified the
+// adaptive sync strategy.
+func (s *memorySeriesStorage) calculatePersistenceUrgencyScore() float64 {
+	var (
+		chunksToPersist    = float64(s.getNumChunksToPersist())
+		maxChunksToPersist = float64(s.maxChunksToPersist)
+		memChunks          = float64(atomic.LoadInt64(&numMemChunks))
+		maxMemChunks       = float64(s.maxMemoryChunks)
+	)
+	score := chunksToPersist / maxChunksToPersist
+	if chunksToPersist > maxChunksToPersist*factorMinChunksToPersist {
+		score = math.Max(
+			score,
+			(memChunks/maxMemChunks-1)/(toleranceFactorForMemChunks-1),
+		)
 	}
-	s.degraded = nowDegraded
-	return s.degraded
-}
+	if score > 1 {
+		score = 1
+	}
+	s.persistenceUrgencyScore.Set(score)
 
-// persistenceBacklogScore works similar to isDegraded, but returns a score
-// about how close we are to degradation. This score is 1.0 if no chunks are
-// waiting for persistence and 0.0 if we are at or above the degradation
-// threshold.
-func (s *memorySeriesStorage) persistenceBacklogScore() float64 {
-	score := 1 - float64(s.getNumChunksToPersist())/float64(s.maxChunksToPersist*percentChunksToPersistForDegradation/100)
-	if score < 0 {
-		return 0
+	if s.rushed {
+		// We are already in rushed mode. If the score is still above
+		// persintenceUrgencyScoreForLeavingRushedMode, return 1 and
+		// leave things as they are.
+		if score > persintenceUrgencyScoreForLeavingRushedMode {
+			return 1
+		}
+		// We are out of rushed mode!
+		s.rushed = false
+		s.rushedMode.Set(0)
+		log.
+			With("urgencyScore", score).
+			With("chunksToPersist", chunksToPersist).
+			With("maxChunksToPersist", maxChunksToPersist).
+			With("memoryChunks", memChunks).
+			With("maxMemoryChunks", maxMemChunks).
+			Warn("Storage has left rushed mode.")
+		return score
+	}
+	if score > persintenceUrgencyScoreForEnteringRushedMode {
+		// Enter rushed mode.
+		s.rushed = true
+		s.rushedMode.Set(1)
+		log.
+			With("urgencyScore", score).
+			With("chunksToPersist", chunksToPersist).
+			With("maxChunksToPersist", maxChunksToPersist).
+			With("memoryChunks", memChunks).
+			With("maxMemoryChunks", maxMemChunks).
+			Warn("Storage has entered rushed mode.")
+		return 1
 	}
 	return score
 }
@@ -1193,6 +1268,8 @@ func (s *memorySeriesStorage) Describe(ch chan<- *prometheus.Desc) {
 	ch <- s.invalidPreloadRequestsCount.Desc()
 	ch <- numMemChunksDesc
 	s.maintainSeriesDuration.Describe(ch)
+	ch <- s.persistenceUrgencyScore.Desc()
+	ch <- s.rushedMode.Desc()
 }
 
 // Collect implements prometheus.Collector.
@@ -1222,4 +1299,6 @@ func (s *memorySeriesStorage) Collect(ch chan<- prometheus.Metric) {
 		float64(atomic.LoadInt64(&numMemChunks)),
 	)
 	s.maintainSeriesDuration.Collect(ch)
+	ch <- s.persistenceUrgencyScore
+	ch <- s.rushedMode
 }