Fix race condition in custom libbeat instrumentation

changelogs/head.asciidoc

@@ -14,6 +14,7 @@ https://github.com/elastic/apm-server/compare/8.5\...main[View commits]
 ==== Bug fixes
 - Set `message` instead of `labels.event` for Jaeger span events {pull}8765[8765]
 - Fix event loss during reload of TBS processor {pull}8809[8809]
+- Fix sporadically missing custom libbeat metrics {pull}8900[8900]
 
 [float]
 ==== Intake API Changes

internal/beater/beater.go

@@ -235,7 +235,10 @@ func (bt *beater) run(ctx context.Context, cancelContext context.CancelFunc, b *
 		if b.Config != nil {
 			reloader.outputConfig = b.Config.Output
 		}
-		if err := reloader.reload(); err != nil {
+		reloader.mu.Lock()
+		err := reloader.reload()
+		reloader.mu.Unlock()
+		if err != nil {
 			return err
 		}
 	}
@@ -289,6 +292,7 @@ func (r *reloader) Reload(configs []*reload.ConfigWithMeta) error {
 	}
 
 	r.mu.Lock()
+	defer r.mu.Unlock()
 	r.rawConfig = integrationConfig.APMServer
 	// Merge in datastream namespace passed in from apm integration
 	if integrationConfig.DataStream != nil && integrationConfig.DataStream.Namespace != "" {
@@ -300,7 +304,6 @@ func (r *reloader) Reload(configs []*reload.ConfigWithMeta) error {
 		}
 	}
 	r.fleetConfig = &integrationConfig.Fleet
-	r.mu.Unlock()
 	return r.reload()
 }
 
@@ -312,14 +315,16 @@ func (r *reloader) reloadOutput(config *reload.ConfigWithMeta) error {
 		}
 	}
 	r.mu.Lock()
+	defer r.mu.Unlock()
 	r.outputConfig = outputConfig
-	r.mu.Unlock()
 	return r.reload()
 }
 
+// reload creates a new serverRunner, launches it in a new goroutine, waits
+// for it to have successfully started and returns after waiting for the previous
+// serverRunner (if any) to exit. Calls to reload must be sycnhronized explicitly
+// by acquiring reloader#mu by callers.
 func (r *reloader) reload() error {
-	r.mu.Lock()
-	defer r.mu.Unlock()
 	if r.rawConfig == nil {
 		// APM Server config not loaded yet.
 		return nil
@@ -345,6 +350,17 @@ func (r *reloader) reload() error {
 			r.args.Logger.Error(err)
 		}
 	}()
+
+	// Wait for the new runner to start; this avoids the race condition in updating
+	// the monitoring#Deafult global registry inside the runner due to two reloads,
+	// one for the inputs and the other for the elasticsearch output
+	select {
+	case <-runner.done:
+		return errors.New("runner exited unexpectedly")
+	case <-runner.started:
+		// runner has started
+	}
+
 	// If the old runner exists, cancel it
 	if r.runner != nil {
 		r.runner.cancelRunServerContext()
@@ -364,6 +380,7 @@ type serverRunner struct {
 	// immediately when the Stop method is invoked.
 	runServerContext       context.Context
 	cancelRunServerContext context.CancelFunc
+	started                chan struct{}
 	done                   chan struct{}
 
 	pipeline                  beat.PipelineConnector
@@ -416,6 +433,7 @@ func newServerRunner(ctx context.Context, args serverRunnerParams) (*serverRunne
 		runServerContext:       runServerContext,
 		cancelRunServerContext: cancel,
 		done:                   make(chan struct{}),
+		started:                make(chan struct{}),
 
 		config:                    cfg,
 		rawConfig:                 args.RawConfig,
@@ -581,6 +599,10 @@ func (s *serverRunner) run(listener net.Listener) error {
 			NewElasticsearchClient: newElasticsearchClient,
 		})
 	})
+
+	// Signal that the runner has started
+	close(s.started)
+
 	result := g.Wait()
 	if err := closeFinalBatchProcessor(s.backgroundContext); err != nil {
 		result = multierror.Append(result, err)
@@ -664,19 +686,12 @@ func (s *serverRunner) waitReady(ctx context.Context, kibanaClient kibana.Client
 	return waitReady(ctx, s.config.WaitReadyInterval, s.tracer, s.logger, check)
 }
 
-// This mutex must be held when updating the libbeat monitoring registry,
-// as there may be multiple servers running concurrently.
-var monitoringRegistryMu sync.Mutex
-
 // newFinalBatchProcessor returns the final model.BatchProcessor that publishes events,
 // and a cleanup function which should be called on server shutdown. If the output is
 // "elasticsearch", then we use modelindexer; otherwise we use the libbeat publisher.
 func (s *serverRunner) newFinalBatchProcessor(
 	newElasticsearchClient func(cfg *elasticsearch.Config) (elasticsearch.Client, error),
 ) (model.BatchProcessor, func(context.Context) error, error) {
-	monitoringRegistryMu.Lock()
-	defer monitoringRegistryMu.Unlock()
-
 	if s.elasticsearchOutputConfig == nil {
 		// When the publisher stops cleanly it will close its pipeline client,
 		// calling the acker's Close method. We need to call Open for each new