Missing latency by HTTP status code for Knative Serving with Kourier

[gtupak]

Ask your question here:

I'm trying to monitor request latency for my Knative services broken down by HTTP response code (or response code class like 2xx, 4xx, 5xx), similar to how I can monitor RPS.

The Envoy metrics envoy_cluster_external_upstream_rq_time_sum and envoy_cluster_external_upstream_rq_time_count don't include the envoy_response_code or envoy_response_code_class labels, so I cannot break down latency by response status.

I've checked:

• envoy_cluster_external_upstream_rq_time_* - no response code labels
• envoy_cluster_upstream_rq_time_* - no response code labels
• envoy_http_downstream_rq_time_* - no cluster-specific or response code labels
• kn_revision_* metrics - these track autoscaler metrics but not request latency

What is the recommended way to monitor request latency per response code (or response code class) for Knative services?

Environment

• Knative Serving version: [your version]
• Ingress: Kourier (3scale-kourier-gateway)
• Monitoring: Prometheus + Grafana

[dprotaso]

cc @dsimansk do you know what metrics kourier emits?

[linkvt]

Hi,

I think this is rather an envoy issue as envoy sits between client and server. My understanding is that the net-kourier-controller is unaware of details about the requests and can't know anything them.

I also just checked my envoy metrics (using 1.36) and also the metrics of the net-kourier-controller and couldn't find anything that would help, same as you describe in the issue description.

Not sure if there is anything we can do here

[gtupak]

So how do people usually monitor their knative request latencies? What I ended up doing is putting Nginx in front of kourier but this adds an extra hop that I'd like to avoid.

[linkvt]

Envoy (as part of kourier) exposes its default metrics, the latency is provided per service as part of the envoy_cluster_upstream_rq_time_bucket histogram. You can use that to measure the latency, but as envoy doesn't include labels for latency per status code you can't get metric. There is a similar issue in the envoy repo but it was closed 5 years ago, you could give that another bump if it's important to you:

Besides that, istios sidecar pods do have a metric istio_request_duration_milliseconds that also contains the status code:

istio_request_duration_milliseconds_bucket{reporter="destination",source_workload="unknown",source_canonical_service="unknown",source_canonical_revision="latest",source_workload_namespace="unknown",source_principal="unknown",source_app="unknown",source_version="unknown",source_cluster="unknown",destination_workload="httpbin",destination_workload_namespace="default",destination_principal="unknown",destination_app="httpbin",destination_version="",destination_service="httpbin.default.svc.cluster.local",destination_canonical_service="httpbin",destination_canonical_revision="latest",destination_service_name="httpbin",destination_service_namespace="default",destination_cluster="Kubernetes",request_protocol="http",response_code="404",grpc_response_status="",response_flags="-",connection_security_policy="none",le="0.5"} 0

You could try using istio with knative, its also supported.

You could also give distributed tracing a look, IIRC you could also filter your requests there by the response code.