Adding pprof endpoints to metrics port

Author: kfswain

cmd/epp/runner/runner.go

@@ -20,6 +20,7 @@ import (
 	"context"
 	"flag"
 	"fmt"
+	"net/http/pprof"
 
 	"github.com/go-logr/logr"
 	"github.com/prometheus/client_golang/prometheus"
@@ -215,6 +216,11 @@ func (r *Runner) Run(ctx context.Context) error {
 		setupLog.Error(err, "Failed to create controller manager")
 		return err
 	}
+	err = setupPprofHandlers(mgr)
+	if err != nil {
+		setupLog.Error(err, "Failed to setup pprof handlers")
+		return err
+	}
 
 	if len(*configText) != 0 || len(*configFile) != 0 {
 		theConfig, err := loader.LoadConfig([]byte(*configText), *configFile)
@@ -403,3 +409,24 @@ func verifyMetricMapping(mapping backendmetrics.MetricMapping, logger logr.Logge
 		logger.Info("Not scraping metric: LoraRequestInfo")
 	}
 }
+
+// setupPprofHandlers only implements the pre-defined profiles:
+// https://cs.opensource.google/go/go/+/refs/tags/go1.24.4:src/runtime/pprof/pprof.go;l=108
+func setupPprofHandlers(mgr ctrl.Manager) error {
+	var err error
+	profiles := []string{
+		"heap",
+		"goroutine",
+		"allocs",
+		"threadcreate",
+		"block",
+		"mutex",
+	}
+	for _, p := range profiles {
+		err = mgr.AddMetricsServerExtraHandler("/debug/pprof/"+p, pprof.Handler(p))
+		if err != nil {
+			return err
+		}
+	}
+	return nil
+}

site-src/guides/metrics.md

@@ -1,6 +1,6 @@
-# Metrics
+# Metrics & Observability
 
-This guide describes the current state of exposed metrics and how to scrape them.
+This guide describes the current state of exposed metrics and how to scrape them, as well as accessing pprof profiles.
 
 ## Requirements
 
@@ -53,7 +53,7 @@ This guide describes the current state of exposed metrics and how to scrape them
 |:---------------------------|:-----------------|:-------------------------------------------------|:------------------------------------------|:------------|
 | lora_syncer_adapter_status | Gauge            | Status of LoRA adapters (1=loaded, 0=not_loaded) | `adapter_name`=<adapter-id>         | ALPHA       |
 
-## Scrape Metrics
+## Scrape Metrics & Pprof profiles
 
 The metrics endpoints are exposed on different ports by default:
 
@@ -73,6 +73,7 @@ metadata:
 rules:
 - nonResourceURLs:
   - /metrics
+  - /debug/pprof/*
   verbs:
   - get
 ---
@@ -116,6 +117,16 @@ kubectl -n default port-forward inference-gateway-ext-proc-pod-name  9090
 curl -H "Authorization: Bearer $TOKEN" localhost:9090/metrics
 ```
 
+### Pprof profiles
+
+Currently only the [predefined profiles](https://pkg.go.dev/runtime/pprof#Profile) are supported, CPU profiling will require code changes. Assuming the EPP has been port-forwarded as in the above example, to get the PGN display of the `heap` profile simply run:
+
+```
+PROFILE_NAME=heap
+curl -H "Authorization: Bearer $TOKEN" localhost:9090/debug/pprof/$PROFILE_NAME -o profile.out
+go tool pprof -png profile.out
+```
+
 ## Prometheus Alerts
 
 The section instructs how to configure prometheus alerts using collected metrics.