Add aggregation capability to signalfx exporter translations

We need simple aggregation capability to get machine_cpu_cores metric. For now only "count" aggregation method is needed

exporter/signalfxexporter/factory_test.go

@@ -176,7 +176,7 @@ func TestCreateMetricsExporterWithDefaultTranslaitonRules(t *testing.T) {
 
 	// Validate that default translation rules are loaded
 	// Expected values has to be updated once default config changed
-	assert.Equal(t, 7, len(config.TranslationRules))
+	assert.Equal(t, 9, len(config.TranslationRules))
 	assert.Equal(t, translation.ActionRenameDimensionKeys, config.TranslationRules[0].Action)
 	assert.Equal(t, 15, len(config.TranslationRules[0].Mapping))
 }

exporter/signalfxexporter/translation/constants.go

@@ -116,5 +116,14 @@ translation_rules:
     cpu.wait: int
     cpu.softirq: int
     cpu.nice: int
+
+- action: copy_metrics
+  mapping:
+    cpu.idle: machine_cpu_cores
+
+- action: aggregate_metric
+  metric_name: machine_cpu_cores
+  aggregation_method: count
+  dimension_key: host
 `
 )

exporter/signalfxexporter/translation/metricdata_to_signalfxv2.go

@@ -81,23 +81,17 @@ func MetricDataToSignalFxV2(
 	metricTranslator *MetricTranslator,
 	md consumerdata.MetricsData,
 ) (sfxDataPoints []*sfxpb.DataPoint, numDroppedTimeSeries int) {
-	sfxDataPoints, numDroppedTimeSeries = metricDataToSfxDataPoints(logger, md)
-	if metricTranslator != nil {
-		sfxDataPoints = metricTranslator.TranslateDataPoints(logger, sfxDataPoints)
-	}
+	sfxDataPoints, numDroppedTimeSeries = metricDataToSfxDataPoints(logger, metricTranslator, md)
+
 	sanitizeDataPointDimensions(sfxDataPoints)
 	return
 }
 
 func metricDataToSfxDataPoints(
 	logger *zap.Logger,
+	metricTranslator *MetricTranslator,
 	md consumerdata.MetricsData,
 ) (sfxDataPoints []*sfxpb.DataPoint, numDroppedTimeSeries int) {
-
-	// The final number of data points is likely larger than len(md.Metrics)
-	// but at least that is expected.
-	sfxDataPoints = make([]*sfxpb.DataPoint, 0, len(md.Metrics))
-
 	var err error
 
 	// Labels from Node and Resource.
@@ -120,6 +114,13 @@ func metricDataToSfxDataPoints(
 			continue
 		}
 
+		if sfxDataPoints == nil {
+			// Suppose all metrics has roughly similar number of timeseries
+			sfxDataPoints = make([]*sfxpb.DataPoint, 0, len(md.Metrics)*len(metric.Timeseries))
+		}
+
+		metricDataPoints := make([]*sfxpb.DataPoint, 0, len(metric.Timeseries))
+
 		// Build the fixed parts for this metrics from the descriptor.
 		descriptor := metric.MetricDescriptor
 		metricName := descriptor.Name
@@ -157,15 +158,15 @@ func metricDataToSfxDataPoints(
 				switch pv := dp.Value.(type) {
 				case *metricspb.Point_Int64Value:
 					sfxDataPoint.Value = sfxpb.Datum{IntValue: &pv.Int64Value}
-					sfxDataPoints = append(sfxDataPoints, sfxDataPoint)
+					metricDataPoints = append(metricDataPoints, sfxDataPoint)
 
 				case *metricspb.Point_DoubleValue:
 					sfxDataPoint.Value = sfxpb.Datum{DoubleValue: &pv.DoubleValue}
-					sfxDataPoints = append(sfxDataPoints, sfxDataPoint)
+					metricDataPoints = append(metricDataPoints, sfxDataPoint)
 
 				case *metricspb.Point_DistributionValue:
-					sfxDataPoints, err = appendDistributionValues(
-						sfxDataPoints,
+					metricDataPoints, err = appendDistributionValues(
+						metricDataPoints,
 						sfxDataPoint,
 						pv.DistributionValue)
 					if err != nil {
@@ -176,8 +177,8 @@ func metricDataToSfxDataPoints(
 							zap.String("metric", sfxDataPoint.Metric))
 					}
 				case *metricspb.Point_SummaryValue:
-					sfxDataPoints, err = appendSummaryValues(
-						sfxDataPoints,
+					metricDataPoints, err = appendSummaryValues(
+						metricDataPoints,
 						sfxDataPoint,
 						pv.SummaryValue)
 					if err != nil {
@@ -196,6 +197,12 @@ func metricDataToSfxDataPoints(
 
 			}
 		}
+
+		if metricTranslator != nil {
+			metricDataPoints = metricTranslator.TranslateDataPoints(logger, metricDataPoints)
+		}
+
+		sfxDataPoints = append(sfxDataPoints, metricDataPoints...)
 	}
 
 	return sfxDataPoints, numDroppedTimeSeries

exporter/signalfxexporter/translation/translator.go

@@ -65,6 +65,28 @@ const (
 	// k8s.pod.network.io{direction="receive"} -> pod_network_receive_bytes_total{}
 	// k8s.pod.network.io{direction="transmit"} -> pod_network_transmit_bytes_total{}
 	ActionSplitMetric Action = "split_metric"
+
+	// ActionAggregateMetric aggregates metrics by one dimension provided in tr.DimensionKey.
+	// It takes datapoints with name tr.MetricName and aggregates them to a smaller set keeping the same name.
+	// It drops all other dimensions if they have different values across aggregated datapoints,
+	// dimensions with the same values across aggregated datapoints are kept.
+	// tr.DimensionKey is always kept because it is used to aggregate datapoints by.
+	// If a datapoint doesn't have a dimension with tr.DimensionKey, it will be dropped.
+	// tr.AggregationMethod is used to specify a method to aggregate the values.
+	// For example, having the following translation rule:
+	// - action: aggregate_metric
+	//   metric_name: machine_cpu_cores
+	//   aggregation_method: count
+	//   dimension_key: host
+	// The following translations will be performed:
+	// Original datapoints:
+	//   machine_cpu_cores{cpu="cpu1",host="host1"} 0.22
+	//   machine_cpu_cores{cpu="cpu2",host="host1"} 0.11
+	//   machine_cpu_cores{cpu="cpu1",host="host2"} 0.33
+	// Transformed datapoints:
+	//   machine_cpu_cores{host="host1"} 2
+	//   machine_cpu_cores{host="host2"} 1
+	ActionAggregateMetric Action = "aggregate_metric"
 )
 
 // MetricValueType is the enum to capture valid metric value types that can be converted
@@ -77,6 +99,14 @@ const (
 	MetricValueTypeDouble MetricValueType = "double"
 )
 
+// AggregationMethod is the enum used to capture aggregation method
+type AggregationMethod string
+
+const (
+	// AggregationMethodCount represents count aggregation method
+	AggregationMethodCount AggregationMethod = "count"
+)
+
 type Rule struct {
 	// Action specifies the translation action to be applied on metrics.
 	// This is a required field.
@@ -105,6 +135,8 @@ type Rule struct {
 	// TypesMapping is represents metric_name/metric_type key/value pairs,
 	// used by ActionConvertValues.
 	TypesMapping map[string]MetricValueType `mapstructure:"types_mapping"`
+
+	AggregationMethod AggregationMethod `mapstructure:"aggregation_method"`
 }
 
 type MetricTranslator struct {
@@ -178,6 +210,16 @@ func validateTranslationRules(rules []Rule) error {
 					return fmt.Errorf("invalid value type %q set for metric %q in \"types_mapping\"", v, k)
 				}
 			}
+		case ActionAggregateMetric:
+			if tr.MetricName == "" || tr.DimensionKey == "" || tr.AggregationMethod == "" {
+				return fmt.Errorf("fields \"metric_name\", \"dimension_key\", and \"aggregation_method\" "+
+					"are required for %q translation rule", tr.Action)
+			}
+			if tr.AggregationMethod != "count" {
+				return fmt.Errorf("invalid \"aggregation_method\": %q provided for %q translation rule",
+					tr.AggregationMethod, tr.Action)
+			}
+
 		default:
 			return fmt.Errorf("unknown \"action\" value: %q", tr.Action)
 		}
@@ -197,63 +239,99 @@ func createDimensionsMap(rules []Rule) map[string]string {
 	return nil
 }
 
+// TranslateDataPoints transforms datapoints to a format compatible with signalfx backend
+// sfxDataPoints represents one metric converted to signalfx protobuf datapoints
 func (mp *MetricTranslator) TranslateDataPoints(logger *zap.Logger, sfxDataPoints []*sfxpb.DataPoint) []*sfxpb.DataPoint {
 	processedDataPoints := sfxDataPoints
 
 	for _, tr := range mp.rules {
-		newDataPoints := []*sfxpb.DataPoint{}
-
-		for _, dp := range processedDataPoints {
-			switch tr.Action {
-			case ActionRenameDimensionKeys:
+		switch tr.Action {
+		case ActionRenameDimensionKeys:
+			for _, dp := range processedDataPoints {
 				for _, d := range dp.Dimensions {
 					if newKey, ok := tr.Mapping[d.Key]; ok {
 						d.Key = newKey
 					}
 				}
-			case ActionRenameMetrics:
+			}
+		case ActionRenameMetrics:
+			for _, dp := range processedDataPoints {
 				if newKey, ok := tr.Mapping[dp.Metric]; ok {
 					dp.Metric = newKey
 				}
-			case ActionMultiplyInt:
+			}
+		case ActionMultiplyInt:
+			for _, dp := range processedDataPoints {
 				if multiplier, ok := tr.ScaleFactorsInt[dp.Metric]; ok {
 					v := dp.GetValue().IntValue
 					if v != nil {
 						*v = *v * multiplier
 					}
 				}
-			case ActionDivideInt:
+			}
+		case ActionDivideInt:
+			for _, dp := range processedDataPoints {
 				if divisor, ok := tr.ScaleFactorsInt[dp.Metric]; ok {
 					v := dp.GetValue().IntValue
 					if v != nil {
 						*v = *v / divisor
 					}
 				}
-			case ActionMultiplyFloat:
+			}
+		case ActionMultiplyFloat:
+			for _, dp := range processedDataPoints {
 				if multiplier, ok := tr.ScaleFactorsFloat[dp.Metric]; ok {
 					v := dp.GetValue().DoubleValue
 					if v != nil {
 						*v = *v * multiplier
 					}
 				}
-			case ActionCopyMetrics:
+			}
+		case ActionCopyMetrics:
+			newDataPoints := []*sfxpb.DataPoint{}
+			for _, dp := range processedDataPoints {
 				if newMetric, ok := tr.Mapping[dp.Metric]; ok {
 					newDataPoint := proto.Clone(dp).(*sfxpb.DataPoint)
 					newDataPoint.Metric = newMetric
 					newDataPoints = append(newDataPoints, newDataPoint)
 				}
-			case ActionSplitMetric:
+			}
+			processedDataPoints = append(processedDataPoints, newDataPoints...)
+		case ActionSplitMetric:
+			for _, dp := range processedDataPoints {
 				if tr.MetricName == dp.Metric {
 					splitMetric(dp, tr.DimensionKey, tr.Mapping)
 				}
-			case ActionConvertValues:
+			}
+		case ActionConvertValues:
+			for _, dp := range processedDataPoints {
 				if newType, ok := tr.TypesMapping[dp.Metric]; ok {
 					convertMetricValue(logger, dp, newType)
 				}
 			}
+		case ActionAggregateMetric:
+			// NOTE: Based on the usage of TranslateDataPoints we can assume that the datapoints batch []*sfxpb.DataPoint
+			// represents only one metric and all the datapoints can be aggregated together.
+			var dpsToAggregate []*sfxpb.DataPoint
+			var otherDps []*sfxpb.DataPoint
+			for i, dp := range processedDataPoints {
+				if dp.Metric == tr.MetricName {
+					if dpsToAggregate == nil {
+						dpsToAggregate = make([]*sfxpb.DataPoint, 0, len(processedDataPoints)-i)
+					}
+					dpsToAggregate = append(dpsToAggregate, dp)
+				} else {
+					if otherDps == nil {
+						otherDps = make([]*sfxpb.DataPoint, 0, len(processedDataPoints)-i)
+					}
+					// This slice can contain additional datapoints from a different metric
+					// for example copied in a translation step before
+					otherDps = append(otherDps, dp)
+				}
+			}
+			aggregatedDps := aggregateDatapoints(logger, dpsToAggregate, tr.DimensionKey, tr.AggregationMethod)
+			processedDataPoints = append(otherDps, aggregatedDps...)
 		}
-
-		processedDataPoints = append(processedDataPoints, newDataPoints...)
 	}
 
 	return processedDataPoints
@@ -266,6 +344,87 @@ func (mp *MetricTranslator) TranslateDimension(orig string) string {
 	return orig
 }
 
+// aggregateDatapoints aggregates datapoints assuming that they have
+// the same Timestamp, MetricType, Metric and Source fields.
+func aggregateDatapoints(
+	logger *zap.Logger,
+	dps []*sfxpb.DataPoint,
+	dimensionKey string,
+	aggregation AggregationMethod,
+) []*sfxpb.DataPoint {
+	if len(dps) == 0 {
+		return nil
+	}
+
+	// group datapoints by dimension values
+	dimValToDps := make(map[string][]*sfxpb.DataPoint, len(dps))
+	for i, dp := range dps {
+		var dimensionFound bool
+		for _, d := range dp.Dimensions {
+			if d.Key == dimensionKey {
+				if _, ok := dimValToDps[d.Value]; !ok {
+					// set slice capacity to the possible maximum = len(dps)-i to avoid reallocations
+					dimValToDps[d.Value] = make([]*sfxpb.DataPoint, 0, len(dps)-i)
+				}
+				dimValToDps[d.Value] = append(dimValToDps[d.Value], dp)
+				dimensionFound = true
+			}
+		}
+		if !dimensionFound {
+			logger.Debug("dimension to aggregate by is not found, datapoint is dropped",
+				zap.String("metric", dp.Metric), zap.String("dimension", dimensionKey))
+		}
+	}
+
+	// Get aggregated results
+	result := make([]*sfxpb.DataPoint, 0, len(dimValToDps))
+	for _, dps := range dimValToDps {
+		dp := proto.Clone(dps[0]).(*sfxpb.DataPoint)
+		dp.Dimensions = getCommonDimensions(dps)
+		switch aggregation {
+		case AggregationMethodCount:
+			gauge := sfxpb.MetricType_GAUGE
+			dp.MetricType = &gauge
+			value := int64(len(dps))
+			dp.Value = sfxpb.Datum{
+				IntValue: &value,
+			}
+		}
+		result = append(result, dp)
+	}
+
+	return result
+}
+
+// getCommonDimensions returns list of dimension that are the same across provided datapoints
+func getCommonDimensions(dps []*sfxpb.DataPoint) []*sfxpb.Dimension {
+	if len(dps) == 0 {
+		return nil
+	}
+
+	commonDimensions := make([]*sfxpb.Dimension, len(dps[0].Dimensions))
+	copy(commonDimensions, dps[0].Dimensions)
+
+	for _, dp := range dps[1:] {
+		dimensions := commonDimensions
+		commonDimensions = make([]*sfxpb.Dimension, 0, len(dimensions))
+		for _, cd := range dimensions {
+			var match bool
+			for _, d := range dp.Dimensions {
+				if d.Key == cd.Key && d.Value == cd.Value {
+					match = true
+					break
+				}
+			}
+			if match {
+				commonDimensions = append(commonDimensions, cd)
+			}
+		}
+	}
+
+	return commonDimensions
+}
+
 // splitMetric renames a metric with "dimension key" == dimensionKey to mapping["dimension value"],
 // datapoint not changed if not dimension found equal to dimensionKey:mapping->key.
 func splitMetric(dp *sfxpb.DataPoint, dimensionKey string, mapping map[string]string) {