Fix race condition on error returned after context cancellation (elas…

…tic#12743)

After elastic#11981 docker client watching event uses the context provided by
the metricset wrapper. This produces a race condition on error
reporting: the errors channel will receive a context cancelled error when
the context is done, so both paths can be chosen by the select. If the
errors one is chosen, an error will be reported and a reconnect will be
attempted, that will fail.

Alternativelly we could have created another context and cancel it after
the reporter is done, in a similar fashion to what was done before elastic#11981,
but then we would be breaking the chain of derived contexts.

metricbeat/module/docker/event/event.go

@@ -94,13 +94,21 @@ func (m *MetricSet) Run(ctx context.Context, reporter mb.ReporterV2) {
 				m.reportEvent(reporter, event)
 
 			case err := <-errors:
+				// An error can be received on context cancellation, don't reconnect
+				// if context is done.
+				select {
+				case <-ctx.Done():
+					m.logger.Debug("docker", "Event watcher stopped")
+					return
+				default:
+				}
 				// Restart watch call
-				m.logger.Error("Error watching for docker events: %v", err)
+				m.logger.Errorf("Error watching for docker events: %v", err)
 				time.Sleep(1 * time.Second)
 				break WATCH
 
 			case <-ctx.Done():
-				m.logger.Debug("docker", "event watcher stopped")
+				m.logger.Debug("docker", "Event watcher stopped")
 				return
 			}
 		}