app/vmselect/promql: parallelize incremental aggregation to multiple CPU cores

This may reduce response times for aggregation over big number of time series
with small step between output data points.

app/vmselect/netstorage/netstorage.go

@@ -68,9 +68,10 @@ func (rss *Results) Cancel() {
 // RunParallel runs in parallel f for all the results from rss.
 //
 // f shouldn't hold references to rs after returning.
+// workerID is the id of the worker goroutine that calls f.
 //
 // rss becomes unusable after the call to RunParallel.
-func (rss *Results) RunParallel(f func(rs *Result)) error {
+func (rss *Results) RunParallel(f func(rs *Result, workerID uint)) error {
 	defer func() {
 		putTmpBlocksFile(rss.tbf)
 		rss.tbf = nil
@@ -88,7 +89,7 @@ func (rss *Results) RunParallel(f func(rs *Result)) error {
 
 	// Start workers.
 	for i := 0; i < workersCount; i++ {
-		go func() {
+		go func(workerID uint) {
 			rs := getResult()
 			defer putResult(rs)
 			maxWorkersCount := gomaxprocs / workersCount
@@ -106,13 +107,13 @@ func (rss *Results) RunParallel(f func(rs *Result)) error {
 					// Skip empty blocks.
 					continue
 				}
-				f(rs)
+				f(rs, workerID)
 			}
 			// Drain the remaining work
 			for range workCh {
 			}
 			doneCh <- err
-		}()
+		}(uint(i))
 	}
 
 	// Feed workers with work.

app/vmselect/prometheus/prometheus.go

@@ -83,7 +83,7 @@ func FederateHandler(at *auth.Token, w http.ResponseWriter, r *http.Request) err
 	resultsCh := make(chan *quicktemplate.ByteBuffer)
 	doneCh := make(chan error)
 	go func() {
-		err := rss.RunParallel(func(rs *netstorage.Result) {
+		err := rss.RunParallel(func(rs *netstorage.Result, workerID uint) {
 			bb := quicktemplate.AcquireByteBuffer()
 			WriteFederate(bb, rs)
 			resultsCh <- bb
@@ -181,7 +181,7 @@ func exportHandler(at *auth.Token, w http.ResponseWriter, matches []string, star
 	resultsCh := make(chan *quicktemplate.ByteBuffer, runtime.GOMAXPROCS(-1))
 	doneCh := make(chan error)
 	go func() {
-		err := rss.RunParallel(func(rs *netstorage.Result) {
+		err := rss.RunParallel(func(rs *netstorage.Result, workerID uint) {
 			bb := quicktemplate.AcquireByteBuffer()
 			writeLineFunc(bb, rs)
 			resultsCh <- bb
@@ -413,7 +413,7 @@ func SeriesHandler(at *auth.Token, w http.ResponseWriter, r *http.Request) error
 	resultsCh := make(chan *quicktemplate.ByteBuffer)
 	doneCh := make(chan error)
 	go func() {
-		err := rss.RunParallel(func(rs *netstorage.Result) {
+		err := rss.RunParallel(func(rs *netstorage.Result, workerID uint) {
 			bb := quicktemplate.AcquireByteBuffer()
 			writemetricNameObject(bb, &rs.MetricName)
 			resultsCh <- bb

app/vmselect/promql/aggr_incremental.go

@@ -13,30 +13,37 @@ import (
 var incrementalAggrFuncCallbacksMap = map[string]*incrementalAggrFuncCallbacks{
 	"sum": {
 		updateAggrFunc:   updateAggrSum,
+		mergeAggrFunc:    mergeAggrSum,
 		finalizeAggrFunc: finalizeAggrCommon,
 	},
 	"min": {
 		updateAggrFunc:   updateAggrMin,
+		mergeAggrFunc:    mergeAggrMin,
 		finalizeAggrFunc: finalizeAggrCommon,
 	},
 	"max": {
 		updateAggrFunc:   updateAggrMax,
+		mergeAggrFunc:    mergeAggrMax,
 		finalizeAggrFunc: finalizeAggrCommon,
 	},
 	"avg": {
 		updateAggrFunc:   updateAggrAvg,
+		mergeAggrFunc:    mergeAggrAvg,
 		finalizeAggrFunc: finalizeAggrAvg,
 	},
 	"count": {
 		updateAggrFunc:   updateAggrCount,
+		mergeAggrFunc:    mergeAggrCount,
 		finalizeAggrFunc: finalizeAggrCount,
 	},
 	"sum2": {
 		updateAggrFunc:   updateAggrSum2,
+		mergeAggrFunc:    mergeAggrSum2,
 		finalizeAggrFunc: finalizeAggrCommon,
 	},
 	"geomean": {
 		updateAggrFunc:   updateAggrGeomean,
+		mergeAggrFunc:    mergeAggrGeomean,
 		finalizeAggrFunc: finalizeAggrGeomean,
 	},
 }
@@ -44,8 +51,8 @@ var incrementalAggrFuncCallbacksMap = map[string]*incrementalAggrFuncCallbacks{
 type incrementalAggrFuncContext struct {
 	ae *aggrFuncExpr
 
-	mu sync.Mutex
-	m  map[string]*incrementalAggrContext
+	mLock sync.Mutex
+	m     map[uint]map[string]*incrementalAggrContext
 
 	callbacks *incrementalAggrFuncCallbacks
 }
@@ -53,17 +60,24 @@ type incrementalAggrFuncContext struct {
 func newIncrementalAggrFuncContext(ae *aggrFuncExpr, callbacks *incrementalAggrFuncCallbacks) *incrementalAggrFuncContext {
 	return &incrementalAggrFuncContext{
 		ae:        ae,
-		m:         make(map[string]*incrementalAggrContext, 1),
+		m:         make(map[uint]map[string]*incrementalAggrContext),
 		callbacks: callbacks,
 	}
 }
 
-func (iafc *incrementalAggrFuncContext) updateTimeseries(ts *timeseries) {
+func (iafc *incrementalAggrFuncContext) updateTimeseries(ts *timeseries, workerID uint) {
+	iafc.mLock.Lock()
+	m := iafc.m[workerID]
+	if m == nil {
+		m = make(map[string]*incrementalAggrContext, 1)
+		iafc.m[workerID] = m
+	}
+	iafc.mLock.Unlock()
+
 	removeGroupTags(&ts.MetricName, &iafc.ae.Modifier)
 	bb := bbPool.Get()
 	bb.B = marshalMetricNameSorted(bb.B[:0], &ts.MetricName)
-	iafc.mu.Lock()
-	iac := iafc.m[string(bb.B)]
+	iac := m[string(bb.B)]
 	if iac == nil {
 		tsAggr := &timeseries{
 			Values:     make([]float64, len(ts.Values)),
@@ -75,19 +89,30 @@ func (iafc *incrementalAggrFuncContext) updateTimeseries(ts *timeseries) {
 			ts:     tsAggr,
 			values: make([]float64, len(ts.Values)),
 		}
-		iafc.m[string(bb.B)] = iac
+		m[string(bb.B)] = iac
 	}
-	iafc.callbacks.updateAggrFunc(iac, ts.Values)
-	iafc.mu.Unlock()
 	bbPool.Put(bb)
+	iafc.callbacks.updateAggrFunc(iac, ts.Values)
 }
 
 func (iafc *incrementalAggrFuncContext) finalizeTimeseries() []*timeseries {
-	// There is no need in iafc.mu.Lock here, since getTimeseries must be called
+	// There is no need in iafc.mLock.Lock here, since finalizeTimeseries must be called
 	// without concurrent goroutines touching iafc.
-	tss := make([]*timeseries, 0, len(iafc.m))
+	mGlobal := make(map[string]*incrementalAggrContext)
+	mergeAggrFunc := iafc.callbacks.mergeAggrFunc
+	for _, m := range iafc.m {
+		for k, iac := range m {
+			iacGlobal := mGlobal[k]
+			if iacGlobal == nil {
+				mGlobal[k] = iac
+				continue
+			}
+			mergeAggrFunc(iacGlobal, iac)
+		}
+	}
+	tss := make([]*timeseries, 0, len(mGlobal))
 	finalizeAggrFunc := iafc.callbacks.finalizeAggrFunc
-	for _, iac := range iafc.m {
+	for _, iac := range mGlobal {
 		finalizeAggrFunc(iac)
 		tss = append(tss, iac.ts)
 	}
@@ -96,6 +121,7 @@ func (iafc *incrementalAggrFuncContext) finalizeTimeseries() []*timeseries {
 
 type incrementalAggrFuncCallbacks struct {
 	updateAggrFunc   func(iac *incrementalAggrContext, values []float64)
+	mergeAggrFunc    func(dst, src *incrementalAggrContext)
 	finalizeAggrFunc func(iac *incrementalAggrContext)
 }
 
@@ -129,8 +155,33 @@ func updateAggrSum(iac *incrementalAggrContext, values []float64) {
 		if math.IsNaN(v) {
 			continue
 		}
-		dstValues[i] += v
+		if dstCounts[i] == 0 {
+			dstValues[i] = v
 			dstCounts[i] = 1
+			continue
+		}
+		dstValues[i] += v
+	}
+}
+
+func mergeAggrSum(dst, src *incrementalAggrContext) {
+	srcValues := src.ts.Values
+	dstValues := dst.ts.Values
+	srcCounts := src.values
+	dstCounts := dst.values
+	_ = srcCounts[len(srcValues)-1]
+	_ = dstCounts[len(srcValues)-1]
+	_ = dstValues[len(srcValues)-1]
+	for i, v := range srcValues {
+		if srcCounts[i] == 0 {
+			continue
+		}
+		if dstCounts[i] == 0 {
+			dstValues[i] = v
+			dstCounts[i] = 1
+			continue
+		}
+		dstValues[i] += v
 	}
 }
 
@@ -154,6 +205,29 @@ func updateAggrMin(iac *incrementalAggrContext, values []float64) {
 	}
 }
 
+func mergeAggrMin(dst, src *incrementalAggrContext) {
+	srcValues := src.ts.Values
+	dstValues := dst.ts.Values
+	srcCounts := src.values
+	dstCounts := dst.values
+	_ = srcCounts[len(srcValues)-1]
+	_ = dstCounts[len(srcValues)-1]
+	_ = dstValues[len(srcValues)-1]
+	for i, v := range srcValues {
+		if srcCounts[i] == 0 {
+			continue
+		}
+		if dstCounts[i] == 0 {
+			dstValues[i] = v
+			dstCounts[i] = 1
+			continue
+		}
+		if v < dstValues[i] {
+			dstValues[i] = v
+		}
+	}
+}
+
 func updateAggrMax(iac *incrementalAggrContext, values []float64) {
 	dstValues := iac.ts.Values
 	dstCounts := iac.values
@@ -174,6 +248,29 @@ func updateAggrMax(iac *incrementalAggrContext, values []float64) {
 	}
 }
 
+func mergeAggrMax(dst, src *incrementalAggrContext) {
+	srcValues := src.ts.Values
+	dstValues := dst.ts.Values
+	srcCounts := src.values
+	dstCounts := dst.values
+	_ = srcCounts[len(srcValues)-1]
+	_ = dstCounts[len(srcValues)-1]
+	_ = dstValues[len(srcValues)-1]
+	for i, v := range srcValues {
+		if srcCounts[i] == 0 {
+			continue
+		}
+		if dstCounts[i] == 0 {
+			dstValues[i] = v
+			dstCounts[i] = 1
+			continue
+		}
+		if v > dstValues[i] {
+			dstValues[i] = v
+		}
+	}
+}
+
 func updateAggrAvg(iac *incrementalAggrContext, values []float64) {
 	// Do not use `Rapid calculation methods` at https://en.wikipedia.org/wiki/Standard_deviation,
 	// since it is slower and has no obvious benefits in increased precision.
@@ -195,6 +292,28 @@ func updateAggrAvg(iac *incrementalAggrContext, values []float64) {
 	}
 }
 
+func mergeAggrAvg(dst, src *incrementalAggrContext) {
+	srcValues := src.ts.Values
+	dstValues := dst.ts.Values
+	srcCounts := src.values
+	dstCounts := dst.values
+	_ = srcCounts[len(srcValues)-1]
+	_ = dstCounts[len(srcValues)-1]
+	_ = dstValues[len(srcValues)-1]
+	for i, v := range srcValues {
+		if srcCounts[i] == 0 {
+			continue
+		}
+		if dstCounts[i] == 0 {
+			dstValues[i] = v
+			dstCounts[i] = srcCounts[i]
+			continue
+		}
+		dstValues[i] += v
+		dstCounts[i] += srcCounts[i]
+	}
+}
+
 func finalizeAggrAvg(iac *incrementalAggrContext) {
 	dstValues := iac.ts.Values
 	counts := iac.values
@@ -219,6 +338,15 @@ func updateAggrCount(iac *incrementalAggrContext, values []float64) {
 	}
 }
 
+func mergeAggrCount(dst, src *incrementalAggrContext) {
+	srcValues := src.ts.Values
+	dstValues := dst.ts.Values
+	_ = dstValues[len(srcValues)-1]
+	for i, v := range srcValues {
+		dstValues[i] += v
+	}
+}
+
 func finalizeAggrCount(iac *incrementalAggrContext) {
 	// Nothing to do
 }
@@ -232,8 +360,33 @@ func updateAggrSum2(iac *incrementalAggrContext, values []float64) {
 		if math.IsNaN(v) {
 			continue
 		}
-		dstValues[i] += v * v
+		if dstCounts[i] == 0 {
+			dstValues[i] = v * v
 			dstCounts[i] = 1
+			continue
+		}
+		dstValues[i] += v * v
+	}
+}
+
+func mergeAggrSum2(dst, src *incrementalAggrContext) {
+	srcValues := src.ts.Values
+	dstValues := dst.ts.Values
+	srcCounts := src.values
+	dstCounts := dst.values
+	_ = srcCounts[len(srcValues)-1]
+	_ = dstCounts[len(srcValues)-1]
+	_ = dstValues[len(srcValues)-1]
+	for i, v := range srcValues {
+		if srcCounts[i] == 0 {
+			continue
+		}
+		if dstCounts[i] == 0 {
+			dstValues[i] = v
+			dstCounts[i] = 1
+			continue
+		}
+		dstValues[i] += v
 	}
 }
 
@@ -256,6 +409,28 @@ func updateAggrGeomean(iac *incrementalAggrContext, values []float64) {
 	}
 }
 
+func mergeAggrGeomean(dst, src *incrementalAggrContext) {
+	srcValues := src.ts.Values
+	dstValues := dst.ts.Values
+	srcCounts := src.values
+	dstCounts := dst.values
+	_ = srcCounts[len(srcValues)-1]
+	_ = dstCounts[len(srcValues)-1]
+	_ = dstValues[len(srcValues)-1]
+	for i, v := range srcValues {
+		if srcCounts[i] == 0 {
+			continue
+		}
+		if dstCounts[i] == 0 {
+			dstValues[i] = v
+			dstCounts[i] = srcCounts[i]
+			continue
+		}
+		dstValues[i] *= v
+		dstCounts[i] += srcCounts[i]
+	}
+}
+
 func finalizeAggrGeomean(iac *incrementalAggrContext) {
 	dstValues := iac.ts.Values
 	counts := iac.values

app/vmselect/promql/aggr_incremental_test.go

@@ -0,0 +1,187 @@
+package promql
+
+import (
+	"fmt"
+	"math"
+	"reflect"
+	"runtime"
+	"sync"
+	"testing"
+)
+
+func TestIncrementalAggr(t *testing.T) {
+	defaultTimestamps := []int64{100e3, 200e3, 300e3, 400e3}
+	values := [][]float64{
+		{1, nan, 2, nan},
+		{3, nan, nan, 4},
+		{nan, nan, 5, 6},
+		{7, nan, 8, 9},
+		{4, nan, nan, nan},
+		{2, nan, 3, 2},
+		{0, nan, 1, 1},
+	}
+	tssSrc := make([]*timeseries, len(values))
+	for i, vs := range values {
+		ts := &timeseries{
+			Timestamps: defaultTimestamps,
+			Values:     vs,
+		}
+		tssSrc[i] = ts
+	}
+
+	copyTimeseries := func(tssSrc []*timeseries) []*timeseries {
+		tssDst := make([]*timeseries, len(tssSrc))
+		for i, tsSrc := range tssSrc {
+			var tsDst timeseries
+			tsDst.CopyFromShallowTimestamps(tsSrc)
+			tssDst[i] = &tsDst
+		}
+		return tssDst
+	}
+
+	f := func(name string, valuesExpected []float64) {
+		t.Helper()
+		callbacks := getIncrementalAggrFuncCallbacks(name)
+		ae := &aggrFuncExpr{
+			Name: name,
+		}
+		tssExpected := []*timeseries{{
+			Timestamps: defaultTimestamps,
+			Values:     valuesExpected,
+		}}
+		// run the test multiple times to make sure there are no side effects on concurrency
+		for i := 0; i < 10; i++ {
+			iafc := newIncrementalAggrFuncContext(ae, callbacks)
+			tssSrcCopy := copyTimeseries(tssSrc)
+			if err := testIncrementalParallelAggr(iafc, tssSrcCopy, tssExpected); err != nil {
+				t.Fatalf("unexpected error on iteration %d: %s", i, err)
+			}
+		}
+	}
+
+	t.Run("sum", func(t *testing.T) {
+		t.Parallel()
+		valuesExpected := []float64{17, nan, 19, 22}
+		f("sum", valuesExpected)
+	})
+	t.Run("min", func(t *testing.T) {
+		t.Parallel()
+		valuesExpected := []float64{0, nan, 1, 1}
+		f("min", valuesExpected)
+	})
+	t.Run("max", func(t *testing.T) {
+		t.Parallel()
+		valuesExpected := []float64{7, nan, 8, 9}
+		f("max", valuesExpected)
+	})
+	t.Run("avg", func(t *testing.T) {
+		t.Parallel()
+		valuesExpected := []float64{2.8333333333333335, nan, 3.8, 4.4}
+		f("avg", valuesExpected)
+	})
+	t.Run("count", func(t *testing.T) {
+		t.Parallel()
+		valuesExpected := []float64{6, 0, 5, 5}
+		f("count", valuesExpected)
+	})
+	t.Run("sum2", func(t *testing.T) {
+		t.Parallel()
+		valuesExpected := []float64{79, nan, 103, 138}
+		f("sum2", valuesExpected)
+	})
+	t.Run("geomean", func(t *testing.T) {
+		t.Parallel()
+		valuesExpected := []float64{0, nan, 2.9925557394776896, 3.365865436338599}
+		f("geomean", valuesExpected)
+	})
+}
+
+func testIncrementalParallelAggr(iafc *incrementalAggrFuncContext, tssSrc, tssExpected []*timeseries) error {
+	const workersCount = 3
+	tsCh := make(chan *timeseries)
+	var wg sync.WaitGroup
+	wg.Add(workersCount)
+	for i := 0; i < workersCount; i++ {
+		go func(workerID uint) {
+			defer wg.Done()
+			for ts := range tsCh {
+				runtime.Gosched() // allow other goroutines performing the work
+				iafc.updateTimeseries(ts, workerID)
+			}
+		}(uint(i))
+	}
+	for _, ts := range tssSrc {
+		tsCh <- ts
+	}
+	close(tsCh)
+	wg.Wait()
+	tssActual := iafc.finalizeTimeseries()
+	if err := expectTimeseriesEqual(tssActual, tssExpected); err != nil {
+		return fmt.Errorf("%s; tssActual=%v, tssExpected=%v", err, tssActual, tssExpected)
+	}
+	return nil
+}
+
+func expectTimeseriesEqual(actual, expected []*timeseries) error {
+	if len(actual) != len(expected) {
+		return fmt.Errorf("unexpected number of time series; got %d; want %d", len(actual), len(expected))
+	}
+	mActual := timeseriesToMap(actual)
+	mExpected := timeseriesToMap(expected)
+	if len(mActual) != len(mExpected) {
+		return fmt.Errorf("unexpected number of time series after converting to map; got %d; want %d", len(mActual), len(mExpected))
+	}
+	for k, tsExpected := range mExpected {
+		tsActual := mActual[k]
+		if tsActual == nil {
+			return fmt.Errorf("missing time series for key=%q", k)
+		}
+		if err := expectTsEqual(tsActual, tsExpected); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func timeseriesToMap(tss []*timeseries) map[string]*timeseries {
+	m := make(map[string]*timeseries, len(tss))
+	for _, ts := range tss {
+		k := ts.MetricName.Marshal(nil)
+		m[string(k)] = ts
+	}
+	return m
+}
+
+func expectTsEqual(actual, expected *timeseries) error {
+	mnActual := actual.MetricName.Marshal(nil)
+	mnExpected := expected.MetricName.Marshal(nil)
+	if string(mnActual) != string(mnExpected) {
+		return fmt.Errorf("unexpected metric name; got %q; want %q", mnActual, mnExpected)
+	}
+	if !reflect.DeepEqual(actual.Timestamps, expected.Timestamps) {
+		return fmt.Errorf("unexpected timestamps; got %v; want %v", actual.Timestamps, expected.Timestamps)
+	}
+	if err := compareValues(actual.Values, expected.Values); err != nil {
+		return fmt.Errorf("%s; actual %v; expected %v", err, actual.Values, expected.Values)
+	}
+	return nil
+}
+
+func compareValues(vs1, vs2 []float64) error {
+	if len(vs1) != len(vs2) {
+		return fmt.Errorf("unexpected number of values; got %d; want %d", len(vs1), len(vs2))
+	}
+	for i, v1 := range vs1 {
+		v2 := vs2[i]
+		if math.IsNaN(v1) {
+			if !math.IsNaN(v2) {
+				return fmt.Errorf("unexpected value; got %v; want %v", v1, v2)
+			}
+			continue
+		}
+		if v1 != v2 {
+			return fmt.Errorf("unexpected value; got %v; want %v", v1, v2)
+		}
+	}
+	return nil
+}

app/vmselect/promql/eval.go

@@ -663,7 +663,7 @@ func getRollupMemoryLimiter() *memoryLimiter {
 
 func evalRollupWithIncrementalAggregate(iafc *incrementalAggrFuncContext, rss *netstorage.Results, rcs []*rollupConfig,
 	preFunc func(values []float64, timestamps []int64), sharedTimestamps []int64) ([]*timeseries, error) {
-	err := rss.RunParallel(func(rs *netstorage.Result) {
+	err := rss.RunParallel(func(rs *netstorage.Result, workerID uint) {
 		preFunc(rs.Values, rs.Timestamps)
 		ts := getTimeseries()
 		defer putTimeseries(ts)
@@ -675,7 +675,7 @@ func evalRollupWithIncrementalAggregate(iafc *incrementalAggrFuncContext, rss *n
 			}
 			ts.Values = rc.Do(ts.Values[:0], rs.Values, rs.Timestamps)
 			ts.Timestamps = sharedTimestamps
-			iafc.updateTimeseries(ts)
+			iafc.updateTimeseries(ts, workerID)
 			ts.Timestamps = nil
 		}
 	})
@@ -690,7 +690,7 @@ func evalRollupNoIncrementalAggregate(rss *netstorage.Results, rcs []*rollupConf
 	preFunc func(values []float64, timestamps []int64), sharedTimestamps []int64) ([]*timeseries, error) {
 	tss := make([]*timeseries, 0, rss.Len()*len(rcs))
 	var tssLock sync.Mutex
-	err := rss.RunParallel(func(rs *netstorage.Result) {
+	err := rss.RunParallel(func(rs *netstorage.Result, workerID uint) {
 		preFunc(rs.Values, rs.Timestamps)
 		for _, rc := range rcs {
 			var ts timeseries