Make batch processor limit number of data points rather than number of metrics.

This change introduces a BenchmarkBatchMetricProcessor that stress tests batching logic.
Results before:
`BenchmarkBatchMetricProcessor-12           20000             80614 ns/op`
Results after the change:
`BenchmarkBatchMetricProcessor-12           20000             96184 ns/op`

Author: kisieland

CHANGELOG.md

@@ -10,6 +10,7 @@
 - Move BigEndian helper functions in `tracetranslator` to an internal package.(#3298)
 - Rename `configtest.LoadConfigFile` to `configtest.LoadConfigAndValidate` (#3306)
 - Replace `ExtensionCreateParams` with `ExtensionCreateSettings` (#3294)
+- Change `batchprocessor` logic to limit data points rather than metrics (#3141)
 
 ## 💡 Enhancements 💡
 

processor/batchprocessor/README.md

@@ -15,8 +15,8 @@ any data drops such as sampling.
 Please refer to [config.go](./config.go) for the config spec.
 
 The following configuration options can be modified:
-- `send_batch_size` (default = 8192): Number of spans or metrics after which a
-batch will be sent regardless of the timeout.
+- `send_batch_size` (default = 8192): Number of spans, metric data points, or log
+records after which a batch will be sent regardless of the timeout.
 - `timeout` (default = 200ms): Time duration after which a batch will be sent
 regardless of size.
 - `send_batch_max_size` (default = 0): The upper limit of the batch size.

processor/batchprocessor/batch_processor.go

@@ -263,9 +263,9 @@ func (bt *batchTraces) size() int {
 }
 
 type batchMetrics struct {
-	nextConsumer consumer.Metrics
-	metricData   pdata.Metrics
-	metricCount  int
+	nextConsumer   consumer.Metrics
+	metricData     pdata.Metrics
+	dataPointCount int
 }
 
 func newBatchMetrics(nextConsumer consumer.Metrics) *batchMetrics {
@@ -274,19 +274,19 @@ func newBatchMetrics(nextConsumer consumer.Metrics) *batchMetrics {
 
 func (bm *batchMetrics) export(ctx context.Context, sendBatchMaxSize int) error {
 	var req pdata.Metrics
-	if sendBatchMaxSize > 0 && bm.metricCount > sendBatchMaxSize {
+	if sendBatchMaxSize > 0 && bm.dataPointCount > sendBatchMaxSize {
 		req = splitMetrics(sendBatchMaxSize, bm.metricData)
-		bm.metricCount -= sendBatchMaxSize
+		bm.dataPointCount -= sendBatchMaxSize
 	} else {
 		req = bm.metricData
 		bm.metricData = pdata.NewMetrics()
-		bm.metricCount = 0
+		bm.dataPointCount = 0
 	}
 	return bm.nextConsumer.ConsumeMetrics(ctx, req)
 }
 
 func (bm *batchMetrics) itemCount() int {
-	return bm.metricCount
+	return bm.dataPointCount
 }
 
 func (bm *batchMetrics) size() int {
@@ -296,11 +296,11 @@ func (bm *batchMetrics) size() int {
 func (bm *batchMetrics) add(item interface{}) {
 	md := item.(pdata.Metrics)
 
-	newMetricsCount := md.MetricCount()
-	if newMetricsCount == 0 {
+	_, newDataPointCount := md.MetricAndDataPointCount()
+	if newDataPointCount == 0 {
 		return
 	}
-	bm.metricCount += newMetricsCount
+	bm.dataPointCount += newDataPointCount
 	md.ResourceMetrics().MoveAndAppendTo(bm.metricData.ResourceMetrics())
 }
 

processor/batchprocessor/batch_processor_test.go

@@ -17,6 +17,7 @@ package batchprocessor
 import (
 	"context"
 	"fmt"
+	"sync"
 	"testing"
 	"time"
 
@@ -29,6 +30,7 @@ import (
 	"go.opentelemetry.io/collector/component/componenttest"
 	"go.opentelemetry.io/collector/config"
 	"go.opentelemetry.io/collector/config/configtelemetry"
+	"go.opentelemetry.io/collector/consumer"
 	"go.opentelemetry.io/collector/consumer/consumertest"
 	"go.opentelemetry.io/collector/consumer/pdata"
 	"go.opentelemetry.io/collector/internal/testdata"
@@ -320,6 +322,8 @@ func TestBatchMetricProcessor_BatchSize(t *testing.T) {
 
 	requestCount := 100
 	metricsPerRequest := 5
+	dataPointsPerMetric := 2 // Since the int counter uses two datapoints.
+	dataPointsPerRequest := metricsPerRequest * dataPointsPerMetric
 	sink := new(consumertest.MetricsSink)
 
 	createParams := component.ProcessorCreateParams{Logger: zap.NewNop()}
@@ -339,8 +343,8 @@ func TestBatchMetricProcessor_BatchSize(t *testing.T) {
 	elapsed := time.Since(start)
 	require.LessOrEqual(t, elapsed.Nanoseconds(), cfg.Timeout.Nanoseconds())
 
-	expectedBatchesNum := requestCount * metricsPerRequest / int(cfg.SendBatchSize)
-	expectedBatchingFactor := int(cfg.SendBatchSize) / metricsPerRequest
+	expectedBatchesNum := requestCount * dataPointsPerRequest / int(cfg.SendBatchSize)
+	expectedBatchingFactor := int(cfg.SendBatchSize) / dataPointsPerRequest
 
 	require.Equal(t, requestCount*metricsPerRequest, sink.MetricsCount())
 	receivedMds := sink.AllMetrics()
@@ -357,7 +361,7 @@ func TestBatchMetricProcessor_BatchSize(t *testing.T) {
 	assert.Equal(t, 1, len(viewData))
 	distData := viewData[0].Data.(*view.DistributionData)
 	assert.Equal(t, int64(expectedBatchesNum), distData.Count)
-	assert.Equal(t, sink.MetricsCount(), int(distData.Sum()))
+	assert.Equal(t, sink.MetricsCount()*dataPointsPerMetric, int(distData.Sum()))
 	assert.Equal(t, cfg.SendBatchSize, uint32(distData.Min))
 	assert.Equal(t, cfg.SendBatchSize, uint32(distData.Max))
 
@@ -369,6 +373,23 @@ func TestBatchMetricProcessor_BatchSize(t *testing.T) {
 	assert.Equal(t, size, int(distData.Sum()))
 }
 
+func TestBatchMetrics_UnevenBatchMaxSize(t *testing.T) {
+	ctx := context.Background()
+	sink := new(metricsSink)
+	metricsCount := 50
+	dataPointsPerMetric := 2
+	sendBatchMaxSize := 99
+
+	batchMetrics := newBatchMetrics(sink)
+	md := testdata.GenerateMetricsManyMetricsSameResource(metricsCount)
+
+	batchMetrics.add(md)
+	require.Equal(t, dataPointsPerMetric*metricsCount, batchMetrics.dataPointCount)
+	require.NoError(t, batchMetrics.export(ctx, sendBatchMaxSize))
+	remainingDataPointsCount := metricsCount*dataPointsPerMetric - sendBatchMaxSize
+	require.Equal(t, remainingDataPointsCount, batchMetrics.dataPointCount)
+}
+
 func TestBatchMetricsProcessor_Timeout(t *testing.T) {
 	cfg := Config{
 		ProcessorSettings: config.NewProcessorSettings(config.NewID(typeStr)),
@@ -502,6 +523,55 @@ func BenchmarkTraceSizeSpanCount(b *testing.B) {
 	}
 }
 
+func BenchmarkBatchMetricProcessor(b *testing.B) {
+	b.StopTimer()
+	cfg := Config{
+		ProcessorSettings: config.NewProcessorSettings(config.NewID(typeStr)),
+		Timeout:           100 * time.Millisecond,
+		SendBatchSize:     2000,
+	}
+	ctx := context.Background()
+	sink := new(metricsSink)
+	createParams := component.ProcessorCreateParams{Logger: zap.NewNop()}
+	metricsPerRequest := 1000
+
+	batcher, err := newBatchMetricsProcessor(createParams, sink, &cfg, configtelemetry.LevelDetailed)
+	require.NoError(b, err)
+	require.NoError(b, batcher.Start(ctx, componenttest.NewNopHost()))
+
+	mds := make([]pdata.Metrics, 0, b.N)
+	for n := 0; n < b.N; n++ {
+		mds = append(mds,
+			testdata.GenerateMetricsManyMetricsSameResource(metricsPerRequest),
+		)
+	}
+	b.StartTimer()
+	for n := 0; n < b.N; n++ {
+		batcher.ConsumeMetrics(ctx, mds[n])
+	}
+	b.StopTimer()
+	require.NoError(b, batcher.Shutdown(ctx))
+	require.Equal(b, b.N*metricsPerRequest, sink.metricsCount)
+}
+
+type metricsSink struct {
+	mu           sync.Mutex
+	metricsCount int
+}
+
+func (sme *metricsSink) Capabilities() consumer.Capabilities {
+	return consumer.Capabilities{
+		MutatesData: false,
+	}
+}
+
+func (sme *metricsSink) ConsumeMetrics(_ context.Context, md pdata.Metrics) error {
+	sme.mu.Lock()
+	defer sme.mu.Unlock()
+	sme.metricsCount += md.MetricCount()
+	return nil
+}
+
 func TestBatchLogProcessor_ReceivingData(t *testing.T) {
 	// Instantiate the batch processor with low config values to test data
 	// gets sent through the processor.

processor/batchprocessor/splitmetrics.go

@@ -20,15 +20,16 @@ import (
 
 // splitMetrics removes metrics from the input data and returns a new data of the specified size.
 func splitMetrics(size int, src pdata.Metrics) pdata.Metrics {
-	if src.MetricCount() <= size {
+	_, dataPoints := src.MetricAndDataPointCount()
+	if dataPoints <= size {
 		return src
 	}
-	totalCopiedMetrics := 0
+	totalCopiedDataPoints := 0
 	dest := pdata.NewMetrics()
 
 	src.ResourceMetrics().RemoveIf(func(srcRs pdata.ResourceMetrics) bool {
 		// If we are done skip everything else.
-		if totalCopiedMetrics == size {
+		if totalCopiedDataPoints == size {
 			return false
 		}
 
@@ -37,29 +38,31 @@ func splitMetrics(size int, src pdata.Metrics) pdata.Metrics {
 
 		srcRs.InstrumentationLibraryMetrics().RemoveIf(func(srcIlm pdata.InstrumentationLibraryMetrics) bool {
 			// If we are done skip everything else.
-			if totalCopiedMetrics == size {
+			if totalCopiedDataPoints == size {
 				return false
 			}
 
 			destIlm := destRs.InstrumentationLibraryMetrics().AppendEmpty()
 			srcIlm.InstrumentationLibrary().CopyTo(destIlm.InstrumentationLibrary())
 
 			// If possible to move all metrics do that.
-			srcMetricsLen := srcIlm.Metrics().Len()
-			if size-totalCopiedMetrics >= srcMetricsLen {
-				totalCopiedMetrics += srcMetricsLen
+			srcDataPointCount := metricSliceDataPointCount(srcIlm.Metrics())
+			if size-totalCopiedDataPoints >= srcDataPointCount {
+				totalCopiedDataPoints += srcDataPointCount
 				srcIlm.Metrics().MoveAndAppendTo(destIlm.Metrics())
 				return true
 			}
 
 			srcIlm.Metrics().RemoveIf(func(srcMetric pdata.Metric) bool {
 				// If we are done skip everything else.
-				if totalCopiedMetrics == size {
+				if totalCopiedDataPoints == size {
 					return false
 				}
-				srcMetric.CopyTo(destIlm.Metrics().AppendEmpty())
-				totalCopiedMetrics++
-				return true
+				// If the metric has more data points than free slots we should split it.
+				newMetric, remove := splitMetric(srcMetric, size-totalCopiedDataPoints)
+				newMetric.CopyTo(destIlm.Metrics().AppendEmpty())
+				totalCopiedDataPoints += metricDataPointCount(newMetric)
+				return remove
 			})
 			return false
 		})
@@ -68,3 +71,83 @@ func splitMetrics(size int, src pdata.Metrics) pdata.Metrics {
 
 	return dest
 }
+
+// metricSliceDataPointCount calculates the total number of  data points.
+func metricSliceDataPointCount(ms pdata.MetricSlice) (dataPointCount int) {
+	for k := 0; k < ms.Len(); k++ {
+		dataPointCount += metricDataPointCount(ms.At(k))
+	}
+	return
+}
+
+// metricDataPointCount calculates the total number of  data points.
+func metricDataPointCount(ms pdata.Metric) (dataPointCount int) {
+	switch ms.DataType() {
+	case pdata.MetricDataTypeIntGauge:
+		dataPointCount = ms.IntGauge().DataPoints().Len()
+	case pdata.MetricDataTypeDoubleGauge:
+		dataPointCount = ms.DoubleGauge().DataPoints().Len()
+	case pdata.MetricDataTypeIntSum:
+		dataPointCount = ms.IntSum().DataPoints().Len()
+	case pdata.MetricDataTypeDoubleSum:
+		dataPointCount = ms.DoubleSum().DataPoints().Len()
+	case pdata.MetricDataTypeIntHistogram:
+		dataPointCount = ms.IntHistogram().DataPoints().Len()
+	case pdata.MetricDataTypeHistogram:
+		dataPointCount = ms.Histogram().DataPoints().Len()
+	case pdata.MetricDataTypeSummary:
+		dataPointCount = ms.Summary().DataPoints().Len()
+	}
+	return
+}
+
+// splitMetric removes metric points from the input data and returns new data of the specified size
+// and boolean describing, whether the metric should be removed from original slice.
+func splitMetric(ms pdata.Metric, size int) (pdata.Metric, bool) {
+	if metricDataPointCount(ms) <= size {
+		return ms, true
+	}
+
+	result := pdata.NewMetric()
+	ms.CopyTo(result)
+	msSize, i := metricDataPointCount(ms)-size, 0
+	filterDataPoints := func() bool { i++; return i <= msSize }
+	switch ms.DataType() {
+	case pdata.MetricDataTypeIntGauge:
+		result.IntGauge().DataPoints().Resize(size)
+		ms.IntGauge().DataPoints().RemoveIf(func(_ pdata.IntDataPoint) bool {
+			return filterDataPoints()
+		})
+	case pdata.MetricDataTypeDoubleGauge:
+		result.DoubleGauge().DataPoints().Resize(size)
+		ms.DoubleGauge().DataPoints().RemoveIf(func(_ pdata.DoubleDataPoint) bool {
+			return filterDataPoints()
+		})
+	case pdata.MetricDataTypeIntSum:
+		result.IntSum().DataPoints().Resize(size)
+		ms.IntSum().DataPoints().RemoveIf(func(_ pdata.IntDataPoint) bool {
+			return filterDataPoints()
+		})
+	case pdata.MetricDataTypeDoubleSum:
+		result.DoubleSum().DataPoints().Resize(size)
+		ms.DoubleSum().DataPoints().RemoveIf(func(_ pdata.DoubleDataPoint) bool {
+			return filterDataPoints()
+		})
+	case pdata.MetricDataTypeIntHistogram:
+		result.IntHistogram().DataPoints().Resize(size)
+		ms.IntHistogram().DataPoints().RemoveIf(func(_ pdata.IntHistogramDataPoint) bool {
+			return filterDataPoints()
+		})
+	case pdata.MetricDataTypeHistogram:
+		result.Histogram().DataPoints().Resize(size)
+		ms.Histogram().DataPoints().RemoveIf(func(_ pdata.HistogramDataPoint) bool {
+			return filterDataPoints()
+		})
+	case pdata.MetricDataTypeSummary:
+		result.Summary().DataPoints().Resize(size)
+		ms.Summary().DataPoints().RemoveIf(func(_ pdata.SummaryDataPoint) bool {
+			return filterDataPoints()
+		})
+	}
+	return result, false
+}