Merge pull request #83427 from logicalhan/metrics-timeout

Fix double counting issue for request metrics on timeout.

staging/src/k8s.io/apiserver/pkg/endpoints/metrics/metrics.go

@@ -177,7 +177,7 @@ var (
 		},
 		[]string{"group", "version", "kind"},
 	)
-	// Because of volatality of the base metric this is pre-aggregated one. Instead of reporing current usage all the time
+	// Because of volatility of the base metric this is pre-aggregated one. Instead of reporting current usage all the time
 	// it reports maximal usage during the last second.
 	currentInflightRequests = compbasemetrics.NewGaugeVec(
 		&compbasemetrics.GaugeOpts{
@@ -187,6 +187,15 @@ var (
 		},
 		[]string{"requestKind"},
 	)
+
+	requestTerminationsTotal = compbasemetrics.NewCounterVec(
+		&compbasemetrics.CounterOpts{
+			Name:           "apiserver_request_terminations_total",
+			Help:           "Number of requests which apiserver terminated in self-defense.",
+			StabilityLevel: compbasemetrics.ALPHA,
+		},
+		[]string{"verb", "group", "version", "resource", "subresource", "scope", "component", "code"},
+	)
 	kubectlExeRegexp = regexp.MustCompile(`^.*((?i:kubectl\.exe))`)
 
 	metrics = []resettableCollector{
@@ -203,6 +212,7 @@ var (
 		WatchEvents,
 		WatchEventsSizes,
 		currentInflightRequests,
+		requestTerminationsTotal,
 	}
 )
 
@@ -236,10 +246,11 @@ func UpdateInflightRequestMetrics(nonmutating, mutating int) {
 	currentInflightRequests.WithLabelValues(MutatingKind).Set(float64(mutating))
 }
 
-// Record records a single request to the standard metrics endpoints. For use by handlers that perform their own
+// RecordRequestTermination records that the request was terminated early as part of a resource
-// processing. All API paths should use InstrumentRouteFunc implicitly. Use this instead of MonitorRequest if
+// preservation or apiserver self-defense mechanism (e.g. timeouts, maxinflight throttling,
-// you already have a RequestInfo object.
+// proxyHandler errors). RecordRequestTermination should only be called zero or one times
-func Record(req *http.Request, requestInfo *request.RequestInfo, component, contentType string, code int, responseSizeInBytes int, elapsed time.Duration) {
+// per request.
+func RecordRequestTermination(req *http.Request, requestInfo *request.RequestInfo, component string, code int) {
 	if requestInfo == nil {
 		requestInfo = &request.RequestInfo{Verb: req.Method, Path: req.URL.Path}
 	}
@@ -251,9 +262,9 @@ func Record(req *http.Request, requestInfo *request.RequestInfo, component, cont
 	// However, we need to tweak it e.g. to differentiate GET from LIST.
 	verb := canonicalVerb(strings.ToUpper(req.Method), scope)
 	if requestInfo.IsResourceRequest {
-		MonitorRequest(req, verb, requestInfo.APIGroup, requestInfo.APIVersion, requestInfo.Resource, requestInfo.Subresource, scope, component, contentType, code, responseSizeInBytes, elapsed)
+		requestTerminationsTotal.WithLabelValues(cleanVerb(verb, req), requestInfo.APIGroup, requestInfo.APIVersion, requestInfo.Resource, requestInfo.Subresource, scope, component, codeToString(code)).Inc()
 	} else {
-		MonitorRequest(req, verb, "", "", "", requestInfo.Path, scope, component, contentType, code, responseSizeInBytes, elapsed)
+		requestTerminationsTotal.WithLabelValues(cleanVerb(verb, req), "", "", "", requestInfo.Path, scope, component, codeToString(code)).Inc()
 	}
 }
 

staging/src/k8s.io/apiserver/pkg/server/filters/maxinflight.go

@@ -178,7 +178,7 @@ func WithMaxInFlightLimit(
 						}
 					}
 				}
-				metrics.Record(r, requestInfo, metrics.APIServerComponent, "", http.StatusTooManyRequests, 0, 0)
+				metrics.RecordRequestTermination(r, requestInfo, metrics.APIServerComponent, http.StatusTooManyRequests)
 				tooManyRequests(r, w)
 			}
 		}

staging/src/k8s.io/apiserver/pkg/server/filters/timeout.go

@@ -59,7 +59,7 @@ func WithTimeoutForNonLongRunningRequests(handler http.Handler, longRunning apir
 
 		postTimeoutFn := func() {
 			cancel()
-			metrics.Record(req, requestInfo, metrics.APIServerComponent, "", http.StatusGatewayTimeout, 0, 0)
+			metrics.RecordRequestTermination(req, requestInfo, metrics.APIServerComponent, http.StatusGatewayTimeout)
 		}
 		return req, time.After(timeout), postTimeoutFn, apierrors.NewTimeoutError(fmt.Sprintf("request did not complete within %s", timeout), 0)
 	}

staging/src/k8s.io/kube-aggregator/pkg/apiserver/handler_proxy.go

@@ -98,7 +98,7 @@ func proxyError(w http.ResponseWriter, req *http.Request, error string, code int
 		return
 	}
 	// TODO: record long-running request differently? The long-running check func does not necessarily match the one of the aggregated apiserver
-	endpointmetrics.Record(req, info, aggregatorComponent, "", code, 0, 0)
+	endpointmetrics.RecordRequestTermination(req, info, aggregatorComponent, code)
 }
 
 func (r *proxyHandler) ServeHTTP(w http.ResponseWriter, req *http.Request) {