Limit number of goroutines for GetMetricData calls

The CloudWatch client already limits the number of concurrent requests
to GetMetricData API via the `-cloudwatch-concurrency` flag, however
in discovery jobs there's no limit to how many goroutines can be spawn
at the same time and get blocked waiting for the client to be available.

With this change, we use an errgroup to block adding goroutines to
handle GetMetricData partitions until the client is actually available.

pkg/job/discovery.go

@@ -16,6 +16,7 @@ import (
 	"github.com/nerdswords/yet-another-cloudwatch-exporter/pkg/job/maxdimassociator"
 	"github.com/nerdswords/yet-another-cloudwatch-exporter/pkg/logging"
 	"github.com/nerdswords/yet-another-cloudwatch-exporter/pkg/model"
+	"golang.org/x/sync/errgroup"
 )
 
 type resourceAssociator interface {
@@ -32,6 +33,7 @@ func runDiscoveryJob(
 	clientTag tagging.Client,
 	clientCloudwatch cloudwatch.Client,
 	metricsPerQuery int,
+	concurrencyLimit int,
 ) ([]*model.TaggedResource, []*model.CloudwatchData) {
 	logger.Debug("Get tagged resources")
 
@@ -61,33 +63,46 @@ func runDiscoveryJob(
 
 	maxMetricCount := metricsPerQuery
 	length := getMetricDataInputLength(job.Metrics)
-	partition := int(math.Ceil(float64(metricDataLength) / float64(maxMetricCount)))
-	logger.Debug("GetMetricData partitions", "total", partition)
+	partitionSize := int(math.Ceil(float64(metricDataLength) / float64(maxMetricCount)))
+	logger.Debug("GetMetricData partitions", "size", partitionSize)
 
-	var wg sync.WaitGroup
-	wg.Add(partition)
+	g, gCtx := errgroup.WithContext(ctx)
+	g.SetLimit(concurrencyLimit)
 
-	getMetricDataOutput := make([][]*cloudwatch.MetricDataResult, partition)
-	count := 0
+	mu := sync.Mutex{}
+	getMetricDataOutput := make([][]*cloudwatch.MetricDataResult, 0, partitionSize)
 
+	count := 0
 	for i := 0; i < metricDataLength; i += maxMetricCount {
-		go func(i, n int) {
-			defer wg.Done()
-			end := i + maxMetricCount
-			if end > metricDataLength {
-				end = metricDataLength
-			}
-			input := getMetricDatas[i:end]
-			data := clientCloudwatch.GetMetricData(ctx, logger, input, svc.Namespace, length, job.Delay, job.RoundingPeriod)
+		start := i
+		end := i + maxMetricCount
+		if end > metricDataLength {
+			end = metricDataLength
+		}
+		partitionNum := count
+		count++
+
+		g.Go(func() error {
+			logger.Debug("GetMetricData partition", "start", start, "end", end, "partitionNum", partitionNum)
+
+			input := getMetricDatas[start:end]
+			data := clientCloudwatch.GetMetricData(gCtx, logger, input, svc.Namespace, length, job.Delay, job.RoundingPeriod)
 			if data != nil {
-				getMetricDataOutput[n] = data
+				mu.Lock()
+				getMetricDataOutput = append(getMetricDataOutput, data)
+				mu.Unlock()
 			} else {
-				logger.Warn("GetMetricData partition empty result", "partition", n, "start", i, "end", end)
+				logger.Warn("GetMetricData partition empty result", "start", start, "end", end, "partitionNum", partitionNum)
 			}
-		}(i, count)
-		count++
+
+			return nil
+		})
+	}
+
+	if err = g.Wait(); err != nil {
+		logger.Error(err, "GetMetricData work group error")
+		return nil, nil
 	}
-	wg.Wait()
 
 	// Update getMetricDatas slice with values and timestamps from API response.
 	// We iterate through the response MetricDataResults and match the result ID