Merge pull request #2494 from tomwilkie/remote-write-sharding

Dynamically reshard the QueueManager based on observed load.

storage/remote/ewma.go

@@ -0,0 +1,66 @@
+// Copyright 2013 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 remote
+
+import (
+	"sync"
+	"sync/atomic"
+	"time"
+)
+
+// ewmaRate tracks an exponentially weighted moving average of a per-second rate.
+type ewmaRate struct {
+	newEvents int64
+	alpha     float64
+	interval  time.Duration
+	lastRate  float64
+	init      bool
+	mutex     sync.Mutex
+}
+
+func newEWMARate(alpha float64, interval time.Duration) ewmaRate {
+	return ewmaRate{
+		alpha:    alpha,
+		interval: interval,
+	}
+}
+
+// rate returns the per-second rate.
+func (r *ewmaRate) rate() float64 {
+	r.mutex.Lock()
+	defer r.mutex.Unlock()
+	return r.lastRate
+}
+
+// tick assumes to be called every r.interval.
+func (r *ewmaRate) tick() {
+	newEvents := atomic.LoadInt64(&r.newEvents)
+	atomic.AddInt64(&r.newEvents, -newEvents)
+	instantRate := float64(newEvents) / r.interval.Seconds()
+
+	r.mutex.Lock()
+	defer r.mutex.Unlock()
+
+	if r.init {
+		r.lastRate += r.alpha * (instantRate - r.lastRate)
+	} else {
+		r.init = true
+		r.lastRate = instantRate
+	}
+}
+
+// inc counts one event.
+func (r *ewmaRate) incr(incr int64) {
+	atomic.AddInt64(&r.newEvents, incr)
+}

storage/remote/queue_manager.go

@@ -14,6 +14,7 @@
 package remote
 
 import (
+	"math"
 	"sync"
 	"time"
 
@@ -32,13 +33,28 @@ const (
 	subsystem = "remote_storage"
 	queue     = "queue"
 
-	defaultShards            = 10
+	// With a maximum of 1000 shards, assuming an average of 100ms remote write
+	// time and 100 samples per batch, we will be able to push 1M samples/s.
+	defaultMaxShards         = 1000
 	defaultMaxSamplesPerSend = 100
-	// The queue capacity is per shard.
-	defaultQueueCapacity     = 100 * 1024 / defaultShards
+
+	// defaultQueueCapacity is per shard - at 1000 shards, this will buffer
+	// 100M samples.  It is configured to buffer 1000 batches, which at 100ms
+	// per batch is 1:40mins.
+	defaultQueueCapacity     = defaultMaxSamplesPerSend * 1000
 	defaultBatchSendDeadline = 5 * time.Second
-	logRateLimit             = 0.1 // Limit to 1 log event every 10s
-	logBurst                 = 10
+
+	// We track samples in/out and how long pushes take using an Exponentially
+	// Weighted Moving Average.
+	ewmaWeight          = 0.2
+	shardUpdateDuration = 10 * time.Second
+
+	// Allow 30% too many shards before scaling down.
+	shardToleranceFraction = 0.3
+
+	// Limit to 1 log event every 10s
+	logRateLimit = 0.1
+	logBurst     = 10
 )
 
 var (
@@ -97,6 +113,15 @@ var (
 		},
 		[]string{queue},
 	)
+	numShards = prometheus.NewGaugeVec(
+		prometheus.GaugeOpts{
+			Namespace: namespace,
+			Subsystem: subsystem,
+			Name:      "shards",
+			Help:      "The number of shards used for parallel sending to the remote storage.",
+		},
+		[]string{queue},
+	)
 )
 
 func init() {
@@ -106,6 +131,7 @@ func init() {
 	prometheus.MustRegister(sentBatchDuration)
 	prometheus.MustRegister(queueLength)
 	prometheus.MustRegister(queueCapacity)
+	prometheus.MustRegister(numShards)
 }
 
 // StorageClient defines an interface for sending a batch of samples to an
@@ -120,7 +146,7 @@ type StorageClient interface {
 // QueueManagerConfig configures a storage queue.
 type QueueManagerConfig struct {
 	QueueCapacity     int           // Number of samples to buffer per shard before we start dropping them.
-	Shards            int           // Number of shards, i.e. amount of concurrency.
+	MaxShards         int           // Max number of shards, i.e. amount of concurrency.
 	MaxSamplesPerSend int           // Maximum number of samples per send.
 	BatchSendDeadline time.Duration // Maximum time sample will wait in buffer.
 	ExternalLabels    model.LabelSet
@@ -132,11 +158,18 @@ type QueueManagerConfig struct {
 // indicated by the provided StorageClient.
 type QueueManager struct {
 	cfg        QueueManagerConfig
-	shards     []chan *model.Sample
-	wg         sync.WaitGroup
-	done       chan struct{}
 	queueName  string
 	logLimiter *rate.Limiter
+
+	shardsMtx   sync.Mutex
+	shards      *shards
+	numShards   int
+	reshardChan chan int
+	quit        chan struct{}
+	wg          sync.WaitGroup
+
+	samplesIn, samplesOut, samplesOutDuration ewmaRate
+	integralAccumulator                       float64
 }
 
 // NewQueueManager builds a new QueueManager.
@@ -144,8 +177,8 @@ func NewQueueManager(cfg QueueManagerConfig) *QueueManager {
 	if cfg.QueueCapacity == 0 {
 		cfg.QueueCapacity = defaultQueueCapacity
 	}
-	if cfg.Shards == 0 {
-		cfg.Shards = defaultShards
+	if cfg.MaxShards == 0 {
+		cfg.MaxShards = defaultMaxShards
 	}
 	if cfg.MaxSamplesPerSend == 0 {
 		cfg.MaxSamplesPerSend = defaultMaxSamplesPerSend
@@ -154,21 +187,22 @@ func NewQueueManager(cfg QueueManagerConfig) *QueueManager {
 		cfg.BatchSendDeadline = defaultBatchSendDeadline
 	}
 
-	shards := make([]chan *model.Sample, cfg.Shards)
-	for i := 0; i < cfg.Shards; i++ {
-		shards[i] = make(chan *model.Sample, cfg.QueueCapacity)
-	}
-
 	t := &QueueManager{
-		cfg:        cfg,
-		shards:     shards,
-		done:       make(chan struct{}),
-		queueName:  cfg.Client.Name(),
-		logLimiter: rate.NewLimiter(logRateLimit, logBurst),
-	}
+		cfg:         cfg,
+		queueName:   cfg.Client.Name(),
+		logLimiter:  rate.NewLimiter(logRateLimit, logBurst),
+		numShards:   1,
+		reshardChan: make(chan int),
+		quit:        make(chan struct{}),
 
+		samplesIn:          newEWMARate(ewmaWeight, shardUpdateDuration),
+		samplesOut:         newEWMARate(ewmaWeight, shardUpdateDuration),
+		samplesOutDuration: newEWMARate(ewmaWeight, shardUpdateDuration),
+	}
+	t.shards = t.newShards(t.numShards)
+	numShards.WithLabelValues(t.queueName).Set(float64(t.numShards))
 	queueCapacity.WithLabelValues(t.queueName).Set(float64(t.cfg.QueueCapacity))
-	t.wg.Add(cfg.Shards)
+
 	return t
 }
 
@@ -193,13 +227,13 @@ func (t *QueueManager) Append(s *model.Sample) error {
 		return nil
 	}
 
-	fp := snew.Metric.FastFingerprint()
-	shard := uint64(fp) % uint64(t.cfg.Shards)
+	t.shardsMtx.Lock()
+	enqueued := t.shards.enqueue(&snew)
+	t.shardsMtx.Unlock()
 
-	select {
-	case t.shards[shard] <- &snew:
+	if enqueued {
 		queueLength.WithLabelValues(t.queueName).Inc()
-	default:
+	} else {
 		droppedSamplesTotal.WithLabelValues(t.queueName).Inc()
 		if t.logLimiter.Allow() {
 			log.Warn("Remote storage queue full, discarding sample. Multiple subsequent messages of this kind may be suppressed.")
@@ -218,25 +252,186 @@ func (*QueueManager) NeedsThrottling() bool {
 // Start the queue manager sending samples to the remote storage.
 // Does not block.
 func (t *QueueManager) Start() {
-	for i := 0; i < t.cfg.Shards; i++ {
-		go t.runShard(i)
-	}
+	t.wg.Add(2)
+	go t.updateShardsLoop()
+	go t.reshardLoop()
+
+	t.shardsMtx.Lock()
+	defer t.shardsMtx.Unlock()
+	t.shards.start()
 }
 
 // Stop stops sending samples to the remote storage and waits for pending
 // sends to complete.
 func (t *QueueManager) Stop() {
 	log.Infof("Stopping remote storage...")
-	for _, shard := range t.shards {
-		close(shard)
-	}
+	close(t.quit)
 	t.wg.Wait()
+
+	t.shardsMtx.Lock()
+	defer t.shardsMtx.Unlock()
+	t.shards.stop()
 	log.Info("Remote storage stopped.")
 }
 
-func (t *QueueManager) runShard(i int) {
+func (t *QueueManager) updateShardsLoop() {
 	defer t.wg.Done()
-	shard := t.shards[i]
+
+	ticker := time.Tick(shardUpdateDuration)
+	for {
+		select {
+		case <-ticker:
+			t.calculateDesiredShards()
+		case <-t.quit:
+			return
+		}
+	}
+}
+
+func (t *QueueManager) calculateDesiredShards() {
+	t.samplesIn.tick()
+	t.samplesOut.tick()
+	t.samplesOutDuration.tick()
+
+	// We use the number of incoming samples as a prediction of how much work we
+	// will need to do next iteration.  We add to this any pending samples
+	// (received - send) so we can catch up with any backlog. We use the average
+	// outgoing batch latency to work out how many shards we need.
+	var (
+		samplesIn          = t.samplesIn.rate()
+		samplesOut         = t.samplesOut.rate()
+		samplesPending     = samplesIn - samplesOut
+		samplesOutDuration = t.samplesOutDuration.rate()
+	)
+
+	// We use an integral accumulator, like in a PID, to help dampen oscillation.
+	t.integralAccumulator = t.integralAccumulator + (samplesPending * 0.1)
+
+	if samplesOut <= 0 {
+		return
+	}
+
+	var (
+		timePerSample = samplesOutDuration / samplesOut
+		desiredShards = (timePerSample * (samplesIn + samplesPending + t.integralAccumulator)) / float64(time.Second)
+	)
+	log.Debugf("QueueManager.caclulateDesiredShards samplesIn=%f, samplesOut=%f, samplesPending=%f, desiredShards=%f",
+		samplesIn, samplesOut, samplesPending, desiredShards)
+
+	// Changes in the number of shards must be greater than shardToleranceFraction.
+	var (
+		lowerBound = float64(t.numShards) * (1. - shardToleranceFraction)
+		upperBound = float64(t.numShards) * (1. + shardToleranceFraction)
+	)
+	log.Debugf("QueueManager.updateShardsLoop %f <= %f <= %f", lowerBound, desiredShards, upperBound)
+	if lowerBound <= desiredShards && desiredShards <= upperBound {
+		return
+	}
+
+	numShards := int(math.Ceil(desiredShards))
+	if numShards > t.cfg.MaxShards {
+		numShards = t.cfg.MaxShards
+	}
+	if numShards == t.numShards {
+		return
+	}
+
+	// Resharding can take some time, and we want this loop
+	// to stay close to shardUpdateDuration.
+	select {
+	case t.reshardChan <- numShards:
+		log.Infof("Remote storage resharding from %d to %d shards.", t.numShards, numShards)
+		t.numShards = numShards
+	default:
+		log.Infof("Currently resharding, skipping.")
+	}
+}
+
+func (t *QueueManager) reshardLoop() {
+	defer t.wg.Done()
+
+	for {
+		select {
+		case numShards := <-t.reshardChan:
+			t.reshard(numShards)
+		case <-t.quit:
+			return
+		}
+	}
+}
+
+func (t *QueueManager) reshard(n int) {
+	numShards.WithLabelValues(t.queueName).Set(float64(n))
+
+	t.shardsMtx.Lock()
+	newShards := t.newShards(n)
+	oldShards := t.shards
+	t.shards = newShards
+	t.shardsMtx.Unlock()
+
+	oldShards.stop()
+
+	// We start the newShards after we have stopped (the therefore completely
+	// flushed) the oldShards, to guarantee we only every deliver samples in
+	// order.
+	newShards.start()
+}
+
+type shards struct {
+	qm     *QueueManager
+	queues []chan *model.Sample
+	done   chan struct{}
+	wg     sync.WaitGroup
+}
+
+func (t *QueueManager) newShards(numShards int) *shards {
+	queues := make([]chan *model.Sample, numShards)
+	for i := 0; i < numShards; i++ {
+		queues[i] = make(chan *model.Sample, t.cfg.QueueCapacity)
+	}
+	s := &shards{
+		qm:     t,
+		queues: queues,
+		done:   make(chan struct{}),
+	}
+	s.wg.Add(numShards)
+	return s
+}
+
+func (s *shards) len() int {
+	return len(s.queues)
+}
+
+func (s *shards) start() {
+	for i := 0; i < len(s.queues); i++ {
+		go s.runShard(i)
+	}
+}
+
+func (s *shards) stop() {
+	for _, shard := range s.queues {
+		close(shard)
+	}
+	s.wg.Wait()
+}
+
+func (s *shards) enqueue(sample *model.Sample) bool {
+	s.qm.samplesIn.incr(1)
+
+	fp := sample.Metric.FastFingerprint()
+	shard := uint64(fp) % uint64(len(s.queues))
+
+	select {
+	case s.queues[shard] <- sample:
+		return true
+	default:
+		return false
+	}
+}
+
+func (s *shards) runShard(i int) {
+	defer s.wg.Done()
+	queue := s.queues[i]
 
 	// Send batches of at most MaxSamplesPerSend samples to the remote storage.
 	// If we have fewer samples than that, flush them out after a deadline
@@ -245,45 +440,48 @@ func (t *QueueManager) runShard(i int) {
 
 	for {
 		select {
-		case s, ok := <-shard:
+		case sample, ok := <-queue:
 			if !ok {
 				if len(pendingSamples) > 0 {
-					log.Infof("Flushing %d samples to remote storage...", len(pendingSamples))
-					t.sendSamples(pendingSamples)
-					log.Infof("Done flushing.")
+					log.Debugf("Flushing %d samples to remote storage...", len(pendingSamples))
+					s.sendSamples(pendingSamples)
+					log.Debugf("Done flushing.")
 				}
 				return
 			}
 
-			queueLength.WithLabelValues(t.queueName).Dec()
-			pendingSamples = append(pendingSamples, s)
+			queueLength.WithLabelValues(s.qm.queueName).Dec()
+			pendingSamples = append(pendingSamples, sample)
 
-			for len(pendingSamples) >= t.cfg.MaxSamplesPerSend {
-				t.sendSamples(pendingSamples[:t.cfg.MaxSamplesPerSend])
-				pendingSamples = pendingSamples[t.cfg.MaxSamplesPerSend:]
+			for len(pendingSamples) >= s.qm.cfg.MaxSamplesPerSend {
+				s.sendSamples(pendingSamples[:s.qm.cfg.MaxSamplesPerSend])
+				pendingSamples = pendingSamples[s.qm.cfg.MaxSamplesPerSend:]
 			}
-		case <-time.After(t.cfg.BatchSendDeadline):
+		case <-time.After(s.qm.cfg.BatchSendDeadline):
 			if len(pendingSamples) > 0 {
-				t.sendSamples(pendingSamples)
+				s.sendSamples(pendingSamples)
 				pendingSamples = pendingSamples[:0]
 			}
 		}
 	}
 }
 
-func (t *QueueManager) sendSamples(s model.Samples) {
+func (s *shards) sendSamples(samples model.Samples) {
 	// Samples are sent to the remote storage on a best-effort basis. If a
 	// sample isn't sent correctly the first time, it's simply dropped on the
 	// floor.
 	begin := time.Now()
-	err := t.cfg.Client.Store(s)
-	duration := time.Since(begin).Seconds()
+	err := s.qm.cfg.Client.Store(samples)
+	duration := time.Since(begin)
 
 	if err != nil {
-		log.Warnf("error sending %d samples to remote storage: %s", len(s), err)
-		failedSamplesTotal.WithLabelValues(t.queueName).Add(float64(len(s)))
+		log.Warnf("error sending %d samples to remote storage: %s", len(samples), err)
+		failedSamplesTotal.WithLabelValues(s.qm.queueName).Add(float64(len(samples)))
 	} else {
-		sentSamplesTotal.WithLabelValues(t.queueName).Add(float64(len(s)))
+		sentSamplesTotal.WithLabelValues(s.qm.queueName).Add(float64(len(samples)))
 	}
-	sentBatchDuration.WithLabelValues(t.queueName).Observe(duration)
+	sentBatchDuration.WithLabelValues(s.qm.queueName).Observe(duration.Seconds())
+
+	s.qm.samplesOut.incr(int64(len(samples)))
+	s.qm.samplesOutDuration.incr(int64(duration))
 }

storage/remote/queue_manager_test.go

@@ -98,8 +98,8 @@ func TestSampleDelivery(t *testing.T) {
 	c.expectSamples(samples[:len(samples)/2])
 
 	m := NewQueueManager(QueueManagerConfig{
-		Client: c,
-		Shards: 1,
+		Client:    c,
+		MaxShards: 1,
 	})
 
 	// These should be received by the client.
@@ -185,8 +185,10 @@ func (c *TestBlockingStorageClient) Name() string {
 }
 
 func (t *QueueManager) queueLen() int {
+	t.shardsMtx.Lock()
+	defer t.shardsMtx.Unlock()
 	queueLength := 0
-	for _, shard := range t.shards {
+	for _, shard := range t.shards.queues {
 		queueLength += len(shard)
 	}
 	return queueLength
@@ -197,7 +199,7 @@ func TestSpawnNotMoreThanMaxConcurrentSendsGoroutines(t *testing.T) {
 	// `MaxSamplesPerSend*Shards` samples should be consumed by the
 	// per-shard goroutines, and then another `MaxSamplesPerSend`
 	// should be left on the queue.
-	n := defaultMaxSamplesPerSend*defaultShards + defaultMaxSamplesPerSend
+	n := defaultMaxSamplesPerSend*1 + defaultMaxSamplesPerSend
 
 	samples := make(model.Samples, 0, n)
 	for i := 0; i < n; i++ {
@@ -214,6 +216,7 @@ func TestSpawnNotMoreThanMaxConcurrentSendsGoroutines(t *testing.T) {
 	m := NewQueueManager(QueueManagerConfig{
 		Client:        c,
 		QueueCapacity: n,
+		MaxShards:     1,
 	})
 
 	m.Start()
@@ -250,7 +253,7 @@ func TestSpawnNotMoreThanMaxConcurrentSendsGoroutines(t *testing.T) {
 	}
 
 	numCalls := c.NumCalls()
-	if numCalls != uint64(defaultShards) {
-		t.Errorf("Saw %d concurrent sends, expected %d", numCalls, defaultShards)
+	if numCalls != uint64(1) {
+		t.Errorf("Saw %d concurrent sends, expected 1", numCalls)
 	}
 }