Reduce ingester memory by using TSDB's index (#3951)

* Add BenchmarkIngesterV2Push

So we can check the efficiency of changes to `v2Push()`.
Broadly copied from `BenchmarkIngesterPush()`.

Signed-off-by: Bryan Boreham <bjboreham@gmail.com>

* Use TSDB's index of series, remove RefCache

prometheus/prometheus#8600 adds a method to
`TSDB.Appender` which allows us to save building
a parallel cache, reducing ingester heap by about 20%.

We depend on values from GetRef() remaining valid while v2Push() uses
them. Currently the only way a ref can be invalidated is by a head
compaction, which cannot happen while v2Push() holds the append lock.

Signed-off-by: Bryan Boreham <bjboreham@gmail.com>

* New version of GetRef() that returns labels

Now we only need to make a copy if GetRef() returns zero

Note Prometheus update brings in JSON marshalling of model.Duration
prometheus/common#280

Signed-off-by: Bryan Boreham <bjboreham@gmail.com>

* Add comment on use of copiedLabels

Signed-off-by: Bryan Boreham <bjboreham@gmail.com>

* Update to Prometheus main branch

Pinned gRPC and other dependencies changed by the update in Prometheus
to avoid taking so much change on this PR.

Signed-off-by: Bryan Boreham <bjboreham@gmail.com>

CHANGELOG.md

@@ -12,6 +12,7 @@
   * `-alertmanager.cluster.peers` instead of `-cluster.peer`
   * `-alertmanager.cluster.peer-timeout` instead of `-cluster.peer-timeout`
 * [FEATURE] Ruler: added `local` backend support to the ruler storage configuration under the `-ruler-storage.` flag prefix. #3932
+* [ENHANCEMENT] Blocks storage: reduce ingester memory by eliminating series reference cache. #3951
 * [ENHANCEMENT] Ruler: optimized `<prefix>/api/v1/rules` and `<prefix>/api/v1/alerts` when ruler sharding is enabled. #3916
 * [ENHANCEMENT] Ruler: added the following metrics when ruler sharding is enabled: #3916
   * `cortex_ruler_clients`

go.mod

@@ -12,7 +12,7 @@ require (
 	github.com/alecthomas/units v0.0.0-20210208195552-ff826a37aa15
 	github.com/alicebob/miniredis v2.5.0+incompatible
 	github.com/armon/go-metrics v0.3.6
-	github.com/aws/aws-sdk-go v1.37.8
+	github.com/aws/aws-sdk-go v1.38.3
 	github.com/bradfitz/gomemcache v0.0.0-20190913173617-a41fca850d0b
 	github.com/cespare/xxhash v1.1.0
 	github.com/dustin/go-humanize v1.0.0
@@ -44,10 +44,10 @@ require (
 	github.com/opentracing/opentracing-go v1.2.0
 	github.com/pkg/errors v0.9.1
 	github.com/prometheus/alertmanager v0.21.1-0.20210310093010-0f9cab6991e6
-	github.com/prometheus/client_golang v1.9.0
+	github.com/prometheus/client_golang v1.10.0
 	github.com/prometheus/client_model v0.2.0
-	github.com/prometheus/common v0.18.0
-	github.com/prometheus/prometheus v1.8.2-0.20210321183757-31a518faab18
+	github.com/prometheus/common v0.20.0
+	github.com/prometheus/prometheus v1.8.2-0.20210324152458-c7a62b95cea0
 	github.com/segmentio/fasthash v0.0.0-20180216231524-a72b379d632e
 	github.com/sony/gobreaker v0.4.1
 	github.com/spf13/afero v1.2.2
@@ -60,11 +60,11 @@ require (
 	go.etcd.io/etcd/client/v3 v3.5.0-alpha.0.0.20210225194612-fa82d11a958a
 	go.etcd.io/etcd/server/v3 v3.5.0-alpha.0.0.20210225194612-fa82d11a958a
 	go.uber.org/atomic v1.7.0
-	golang.org/x/net v0.0.0-20210119194325-5f4716e94777
-	golang.org/x/sync v0.0.0-20201207232520-09787c993a3a
-	golang.org/x/time v0.0.0-20201208040808-7e3f01d25324
-	google.golang.org/api v0.39.0
-	google.golang.org/grpc v1.34.0
+	golang.org/x/net v0.0.0-20210324051636-2c4c8ecb7826
+	golang.org/x/sync v0.0.0-20210220032951-036812b2e83c
+	golang.org/x/time v0.0.0-20210220033141-f8bda1e9f3ba
+	google.golang.org/api v0.42.0
+	google.golang.org/grpc v1.36.0
 	gopkg.in/yaml.v2 v2.4.0
 	gopkg.in/yaml.v3 v3.0.0-20210107192922-496545a6307b
 	sigs.k8s.io/yaml v1.2.0
@@ -101,3 +101,18 @@ replace github.com/go-openapi/strfmt => github.com/go-openapi/strfmt v0.19.5
 replace github.com/go-openapi/swag => github.com/go-openapi/swag v0.19.9
 
 replace github.com/go-openapi/validate => github.com/go-openapi/validate v0.19.8
+
+// Pin these, which are updated as dependencies in Prometheus; we will take those updates separately and carefully
+replace (
+	github.com/aws/aws-sdk-go => github.com/aws/aws-sdk-go v1.37.8
+	github.com/google/pprof => github.com/google/pprof v0.0.0-20210208152844-1612e9be7af6
+	github.com/miekg/dns => github.com/miekg/dns v1.1.38
+	github.com/prometheus/client_golang => github.com/prometheus/client_golang v1.9.0
+	golang.org/x/crypto => golang.org/x/crypto v0.0.0-20201208171446-5f87f3452ae9
+	golang.org/x/net => golang.org/x/net v0.0.0-20210119194325-5f4716e94777
+	golang.org/x/oauth2 => golang.org/x/oauth2 v0.0.0-20210210192628-66670185b0cd
+	golang.org/x/sync => golang.org/x/sync v0.0.0-20201207232520-09787c993a3a
+	golang.org/x/sys => golang.org/x/sys v0.0.0-20210124154548-22da62e12c0c
+	google.golang.org/api => google.golang.org/api v0.39.0
+	google.golang.org/grpc => google.golang.org/grpc v1.34.0
+)

pkg/ingester/ingester_v2.go

@@ -101,7 +101,6 @@ const (
 type userTSDB struct {
 	db             *tsdb.DB
 	userID         string
-	refCache       *cortex_tsdb.RefCache
 	activeSeries   *ActiveSeries
 	seriesInMetric *metricCounter
 	limiter        *Limiter
@@ -185,7 +184,8 @@ func (u *userTSDB) compactHead(blockDuration int64) error {
 
 	defer u.casState(forceCompacting, active)
 
-	// Ingestion of samples in parallel with forced compaction can lead to overlapping blocks.
+	// Ingestion of samples in parallel with forced compaction can lead to overlapping blocks,
+	// and possible invalidation of the references returned from Appender.GetRef().
 	// So we wait for existing in-flight requests to finish. Future push requests would fail until compaction is over.
 	u.pushesInFlight.Wait()
 
@@ -383,7 +383,6 @@ type TSDBState struct {
 	walReplayTime          prometheus.Histogram
 	appenderAddDuration    prometheus.Histogram
 	appenderCommitDuration prometheus.Histogram
-	refCachePurgeDuration  prometheus.Histogram
 	idleTsdbChecks         *prometheus.CounterVec
 }
 
@@ -435,11 +434,6 @@ func newTSDBState(bucketClient objstore.Bucket, registerer prometheus.Registerer
 			Help:    "The total time it takes for a push request to commit samples appended to TSDB.",
 			Buckets: []float64{.001, .005, .01, .025, .05, .1, .25, .5, 1, 2.5, 5, 10},
 		}),
-		refCachePurgeDuration: promauto.With(registerer).NewHistogram(prometheus.HistogramOpts{
-			Name:    "cortex_ingester_tsdb_refcache_purge_duration_seconds",
-			Help:    "The total time it takes to purge the TSDB series reference cache for a single tenant.",
-			Buckets: prometheus.DefBuckets,
-		}),
 
 		idleTsdbChecks: idleTsdbChecks,
 	}
@@ -619,11 +613,6 @@ func (i *Ingester) updateLoop(ctx context.Context) error {
 	rateUpdateTicker := time.NewTicker(i.cfg.RateUpdatePeriod)
 	defer rateUpdateTicker.Stop()
 
-	// We use an hardcoded value for this ticker because there should be no
-	// real value in customizing it.
-	refCachePurgeTicker := time.NewTicker(5 * time.Minute)
-	defer refCachePurgeTicker.Stop()
-
 	var activeSeriesTickerChan <-chan time.Time
 	if i.cfg.ActiveSeriesMetricsEnabled {
 		t := time.NewTicker(i.cfg.ActiveSeriesMetricsUpdatePeriod)
@@ -646,17 +635,6 @@ func (i *Ingester) updateLoop(ctx context.Context) error {
 				db.ingestedRuleSamples.tick()
 			}
 			i.userStatesMtx.RUnlock()
-		case <-refCachePurgeTicker.C:
-			for _, userID := range i.getTSDBUsers() {
-				userDB := i.getTSDB(userID)
-				if userDB == nil {
-					continue
-				}
-
-				startTime := time.Now()
-				userDB.refCache.Purge(startTime.Add(-cortex_tsdb.DefaultRefCacheTTL))
-				i.TSDBState.refCachePurgeDuration.Observe(time.Since(startTime).Seconds())
-			}
 
 		case <-activeSeriesTickerChan:
 			i.v2UpdateActiveSeries()
@@ -683,6 +661,12 @@ func (i *Ingester) v2UpdateActiveSeries() {
 	}
 }
 
+// GetRef() is an extra method added to TSDB to let Cortex check before calling Add()
+type extendedAppender interface {
+	storage.Appender
+	storage.GetRef
+}
+
 // v2Push adds metrics to a block
 func (i *Ingester) v2Push(ctx context.Context, req *cortexpb.WriteRequest) (*cortexpb.WriteResponse, error) {
 	var firstPartialErr error
@@ -738,13 +722,13 @@ func (i *Ingester) v2Push(ctx context.Context, req *cortexpb.WriteRequest) (*cor
 	)
 
 	// Walk the samples, appending them to the users database
-	app := db.Appender(ctx)
+	app := db.Appender(ctx).(extendedAppender)
 	for _, ts := range req.Timeseries {
-		// Check if we already have a cached reference for this series. Be aware
-		// that even if we have a reference it's not guaranteed to be still valid.
 		// The labels must be sorted (in our case, it's guaranteed a write request
 		// has sorted labels once hit the ingester).
-		cachedRef, copiedLabels, cachedRefExists := db.refCache.Ref(startAppend, cortexpb.FromLabelAdaptersToLabels(ts.Labels))
+
+		// Look up a reference for this series.
+		ref, copiedLabels := app.GetRef(cortexpb.FromLabelAdaptersToLabels(ts.Labels))
 
 		// To find out if any sample was added to this series, we keep old value.
 		oldSucceededSamplesCount := succeededSamplesCount
@@ -753,30 +737,18 @@ func (i *Ingester) v2Push(ctx context.Context, req *cortexpb.WriteRequest) (*cor
 			var err error
 
 			// If the cached reference exists, we try to use it.
-			if cachedRefExists {
-				var ref uint64
-				if ref, err = app.Append(cachedRef, copiedLabels, s.TimestampMs, s.Value); err == nil {
+			if ref != 0 {
+				if _, err = app.Append(ref, copiedLabels, s.TimestampMs, s.Value); err == nil {
 					succeededSamplesCount++
-					// This means the reference changes which means we need to update our cache.
-					if ref != cachedRef {
-						db.refCache.SetRef(startAppend, copiedLabels, ref)
-					}
 					continue
 				}
 
 			} else {
-				var ref uint64
-
-				// Copy the label set because both TSDB and the cache may retain it.
+				// Copy the label set because both TSDB and the active series tracker may retain it.
 				copiedLabels = cortexpb.FromLabelAdaptersToLabelsWithCopy(ts.Labels)
 
+				// Retain the reference in case there are multiple samples for the series.
 				if ref, err = app.Append(0, copiedLabels, s.TimestampMs, s.Value); err == nil {
-					db.refCache.SetRef(startAppend, copiedLabels, ref)
-
-					// Set these in case there are multiple samples for the series.
-					cachedRef = ref
-					cachedRefExists = true
-
 					succeededSamplesCount++
 					continue
 				}
@@ -827,11 +799,8 @@ func (i *Ingester) v2Push(ctx context.Context, req *cortexpb.WriteRequest) (*cor
 
 		if i.cfg.ActiveSeriesMetricsEnabled && succeededSamplesCount > oldSucceededSamplesCount {
 			db.activeSeries.UpdateSeries(cortexpb.FromLabelAdaptersToLabels(ts.Labels), startAppend, func(l labels.Labels) labels.Labels {
-				// If we have already made a copy during this push, no need to create new one.
-				if copiedLabels != nil {
-					return copiedLabels
-				}
-				return cortexpb.CopyLabels(l)
+				// we must already have copied the labels if succeededSamplesCount has been incremented.
+				return copiedLabels
 			})
 		}
 	}
@@ -1435,7 +1404,6 @@ func (i *Ingester) createTSDB(userID string) (*userTSDB, error) {
 
 	userDB := &userTSDB{
 		userID:              userID,
-		refCache:            cortex_tsdb.NewRefCache(),
 		activeSeries:        NewActiveSeries(),
 		seriesInMetric:      newMetricCounter(i.limiter),
 		ingestedAPISamples:  newEWMARate(0.2, i.cfg.RateUpdatePeriod),