Merge pull request #169 from prometheus/muchfasterwal

wal: parallelize sample processing

head.go

@@ -15,6 +15,7 @@ package tsdb
 
 import (
 	"math"
+	"runtime"
 	"sort"
 	"sync"
 	"sync/atomic"
@@ -186,29 +187,19 @@ func NewHead(r prometheus.Registerer, l log.Logger, wal WAL, chunkRange int64) (
 	return h, nil
 }
 
-// ReadWAL initializes the head by consuming the write ahead log.
+// processWALSamples adds a partition of samples it receives to the head and passes
-func (h *Head) ReadWAL() error {
+// them on to other workers.
-	defer h.postings.ensureOrder()
+// Samples before the mint timestamp are discarded.
+func (h *Head) processWALSamples(
+	mint int64,
+	partition, total uint64,
+	input <-chan []RefSample, output chan<- []RefSample,
+) (unknownRefs uint64) {
+	defer close(output)
 
-	r := h.wal.Reader()
+	for samples := range input {
-	mint := h.MinTime()
-
-	// Track number of samples that referenced a series we don't know about
-	// for error reporting.
-	var unknownRefs int
-
-	seriesFunc := func(series []RefSeries) {
-		for _, s := range series {
-			h.getOrCreateWithID(s.Ref, s.Labels.Hash(), s.Labels)
-
-			if h.lastSeriesID < s.Ref {
-				h.lastSeriesID = s.Ref
-			}
-		}
-	}
-	samplesFunc := func(samples []RefSample) {
 		for _, s := range samples {
-			if s.T < mint {
+			if s.T < mint || s.Ref%total != partition {
 				continue
 			}
 			ms := h.series.getByID(s.Ref)
@@ -222,6 +213,63 @@ func (h *Head) ReadWAL() error {
 				h.metrics.chunks.Inc()
 			}
 		}
+		output <- samples
+	}
+	return unknownRefs
+}
+
+// ReadWAL initializes the head by consuming the write ahead log.
+func (h *Head) ReadWAL() error {
+	defer h.postings.ensureOrder()
+
+	r := h.wal.Reader()
+	mint := h.MinTime()
+
+	// Track number of samples that referenced a series we don't know about
+	// for error reporting.
+	var unknownRefs uint64
+
+	// Start workers that each process samples for a partition of the series ID space.
+	// They are connected through a ring of channels which ensures that all sample batches
+	// read from the WAL are processed in order.
+	var (
+		n          = runtime.GOMAXPROCS(0)
+		firstInput = make(chan []RefSample, 300)
+		input      = firstInput
+	)
+	for i := 0; i < n; i++ {
+		output := make(chan []RefSample, 300)
+
+		go func(i int, input <-chan []RefSample, output chan<- []RefSample) {
+			unknown := h.processWALSamples(mint, uint64(i), uint64(n), input, output)
+			atomic.AddUint64(&unknownRefs, unknown)
+		}(i, input, output)
+
+		// The output feeds the next worker goroutine. For the last worker,
+		// it feeds the initial input again to reuse the RefSample slices.
+		input = output
+	}
+
+	// TODO(fabxc): series entries spread between samples can starve the sample workers.
+	// Even with bufferd channels, this can impact startup time with lots of series churn.
+	// We must not pralellize series creation itself but could make the indexing asynchronous.
+	seriesFunc := func(series []RefSeries) {
+		for _, s := range series {
+			h.getOrCreateWithID(s.Ref, s.Labels.Hash(), s.Labels)
+
+			if h.lastSeriesID < s.Ref {
+				h.lastSeriesID = s.Ref
+			}
+		}
+	}
+	samplesFunc := func(samples []RefSample) {
+		var buf []RefSample
+		select {
+		case buf = <-input:
+		default:
+			buf = make([]RefSample, 0, len(samples)*11/10)
+		}
+		firstInput <- append(buf[:0], samples...)
 	}
 	deletesFunc := func(stones []Stone) {
 		for _, s := range stones {
@@ -234,13 +282,18 @@ func (h *Head) ReadWAL() error {
 		}
 	}
 
+	err := r.Read(seriesFunc, samplesFunc, deletesFunc)
+
+	// Signal termination to first worker and wait for last one to close its output channel.
+	close(firstInput)
+	for range input {
+	}
+	if err != nil {
+		return errors.Wrap(err, "consume WAL")
+	}
 	if unknownRefs > 0 {
 		level.Warn(h.logger).Log("msg", "unknown series references in WAL samples", "count", unknownRefs)
 	}
-
-	if err := r.Read(seriesFunc, samplesFunc, deletesFunc); err != nil {
-		return errors.Wrap(err, "consume WAL")
-	}
 	return nil
 }
 
@@ -1168,10 +1221,12 @@ func (s *memSeries) append(t int64, v float64) (success, chunkCreated bool) {
 		c = s.cut(t)
 		chunkCreated = true
 	}
+	numSamples := c.chunk.NumSamples()
+
 	if c.maxTime >= t {
 		return false, chunkCreated
 	}
-	if c.chunk.NumSamples() > samplesPerChunk/4 && t >= s.nextAt {
+	if numSamples > samplesPerChunk/4 && t >= s.nextAt {
 		c = s.cut(t)
 		chunkCreated = true
 	}
@@ -1179,7 +1234,7 @@ func (s *memSeries) append(t int64, v float64) (success, chunkCreated bool) {
 
 	c.maxTime = t
 
-	if c.chunk.NumSamples() == samplesPerChunk/4 {
+	if numSamples == samplesPerChunk/4 {
 		_, maxt := rangeForTimestamp(c.minTime, s.chunkRange)
 		s.nextAt = computeChunkEndTime(c.minTime, c.maxTime, maxt)
 	}

wal.go

@@ -846,7 +846,7 @@ func (r *walReader) Read(
 		deletePool sync.Pool
 	)
 	donec := make(chan struct{})
-	datac := make(chan interface{}, 50)
+	datac := make(chan interface{}, 100)
 
 	go func() {
 		defer close(donec)