Merge pull request #106 from Adityakk9031/fix/issue-82

fix(main): sanitize `remoteConfigURL` to prevent 400 Bad Request on bandwidth reports (#82)

.github/CODEOWNERS

@@ -0,0 +1 @@
+* @oschwartz10612 @miloschwartz

.github/ISSUE_TEMPLATE/1.bug_report.yml

@@ -14,12 +14,13 @@ body:
           label: Environment
           description: Please fill out the relevant details below for your environment.
           value: |
-              - OS Type & Version: (e.g., Ubuntu 22.04)
+              - OS Type & Version:
               - Pangolin Version:
+              - Edition (Community or Enterprise):
               - Gerbil Version:
               - Traefik Version:
               - Newt Version:
-              - Olm Version: (if applicable)
+              - Client Version:
       validations:
           required: true
 

.github/dependabot.yml

@@ -1,40 +1,32 @@
 version: 2
+
 updates:
   - package-ecosystem: "gomod"
     directory: "/"
     schedule:
       interval: "daily"
+    open-pull-requests-limit: 1
     groups:
-      dev-patch-updates:
-        dependency-type: "development"
-        update-types:
-          - "patch"
-      dev-minor-updates:
-        dependency-type: "development"
-        update-types:
-          - "minor"
-      prod-patch-updates:
-        dependency-type: "production"
-        update-types:
-          - "patch"
-      prod-minor-updates:
-        dependency-type: "production"
-        update-types:
-          - "minor"
+      go-dependencies:
+        patterns:
+          - "*"
 
   - package-ecosystem: "docker"
     directory: "/"
     schedule:
       interval: "daily"
+    open-pull-requests-limit: 1
     groups:
-      patch-updates:
-        update-types:
-          - "patch"
-      minor-updates:
-        update-types:
-          - "minor"
+      docker-dependencies:
+        patterns:
+          - "*"
 
   - package-ecosystem: "github-actions"
     directory: "/"
     schedule:
       interval: "weekly"
+    open-pull-requests-limit: 1
+    groups:
+      github-actions-dependencies:
+        patterns:
+          - "*"

.github/workflows/cicd.yml

@@ -36,16 +36,16 @@ jobs:
 
     steps:
       - name: Checkout code
-        uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # v6.0.1
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
       - name: Set up QEMU
-        uses: docker/setup-qemu-action@c7c53464625b32c7a7e944ae62b3e17d2b600130 # v3.7.0
+        uses: docker/setup-qemu-action@96fe6ef7f33517b61c61be40b68a1882f3264fb8 # v4.2.0
 
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@8d2750c68a42422c14e847fe6c8ac0403b4cbd6f # v3.12.0
+        uses: docker/setup-buildx-action@37fe631027851001ddb9b187196cc803df7f5f0e # v4.3.0
 
       - name: Log in to Docker Hub
-        uses: docker/login-action@5e57cd118135c172c3672efd75eb46360885c0ef # v3.6.0
+        uses: docker/login-action@dbcb813823bdd20940b903addbd779551569679f # v4.6.0
         with:
           registry: docker.io
           username: ${{ secrets.DOCKER_HUB_USERNAME }}
@@ -57,9 +57,9 @@ jobs:
         shell: bash
 
       - name: Install Go
-        uses: actions/setup-go@7a3fe6cf4cb3a834922a1244abfce67bcef6a0c5 # v6.2.0
+        uses: actions/setup-go@b7ad1dad31e06c5925ef5d2fc7ad053ef454303e # v7.0.0
         with:
-          go-version: 1.25
+          go-version: 1.26
 
       - name: Update version in main.go
         run: |
@@ -80,7 +80,7 @@ jobs:
         shell: bash
 
       - name: Login in to GHCR
-        uses: docker/login-action@5e57cd118135c172c3672efd75eb46360885c0ef # v3.6.0
+        uses: docker/login-action@dbcb813823bdd20940b903addbd779551569679f # v4.6.0
         with:
           registry: ghcr.io
           username: ${{ github.actor }}
@@ -107,8 +107,9 @@ jobs:
         shell: bash
 
       - name: Install cosign
-        # cosign is used to sign and verify container images (key and keyless)
-        uses: sigstore/cosign-installer@faadad0cce49287aee09b3a48701e75088a2c6ad # v4.0.0
+        uses: sigstore/cosign-installer@6f9f17788090df1f26f669e9d70d6ae9567deba6 # v4.1.2
+        with:
+          cosign-release: v3.0.6
 
       - name: Dual-sign and verify (GHCR & Docker Hub)
         # Sign each image by digest using keyless (OIDC) and key-based signing,
@@ -155,7 +156,7 @@ jobs:
         shell: bash
 
       - name: Upload artifacts from /bin
-        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6.0.0
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
         with:
           name: binaries
           path: bin/

.github/workflows/mirror.yaml

@@ -23,7 +23,7 @@ jobs:
           skopeo --version
 
       - name: Install cosign
-        uses: sigstore/cosign-installer@faadad0cce49287aee09b3a48701e75088a2c6ad # v4.0.0
+        uses: sigstore/cosign-installer@6f9f17788090df1f26f669e9d70d6ae9567deba6 # v4.1.2
 
       - name: Input check
         run: |

.github/workflows/test.yml

@@ -14,12 +14,16 @@ jobs:
     runs-on: amd64-runner
 
     steps:
-      - uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # v6.0.1
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+
+      - name: Read go version
+        id: goversion
+        run: echo "version=$(cat .go-version)" >> $GITHUB_OUTPUT
 
       - name: Set up Go
-        uses: actions/setup-go@7a3fe6cf4cb3a834922a1244abfce67bcef6a0c5 # v6.2.0
+        uses: actions/setup-go@b7ad1dad31e06c5925ef5d2fc7ad053ef454303e # v7.0.0
         with:
-          go-version: 1.25
+          go-version: ${{ steps.goversion.outputs.version }}
 
       - name: Build go
         run: go build

.go-version

@@ -1 +1 @@
-1.25
+1.26

Dockerfile

@@ -1,4 +1,5 @@
-FROM golang:1.25-alpine AS builder
+ARG GO_VERSION=1.26
+FROM golang:${GO_VERSION}-alpine AS builder
 
 # Set the working directory inside the container
 WORKDIR /app
@@ -16,7 +17,7 @@ COPY . .
 RUN CGO_ENABLED=0 GOOS=linux go build -o /gerbil
 
 # Start a new stage from scratch
-FROM alpine:3.23 AS runner
+FROM alpine:3.24 AS runner
 
 RUN apk add --no-cache iptables iproute2
 
@@ -25,4 +26,4 @@ COPY entrypoint.sh /
 
 RUN chmod +x /entrypoint.sh
 ENTRYPOINT ["/entrypoint.sh"]
-CMD ["gerbil"]
+CMD ["gerbil"]

Makefile

@@ -1,4 +1,6 @@
 
+GO_VERSION := $(shell cat .go-version 2>/dev/null)
+
 all: build push
 
 docker-build-release:
@@ -7,10 +9,10 @@ docker-build-release:
 		exit 1; \
 	fi
 	docker buildx build --platform linux/arm64,linux/amd64 -t fosrl/gerbil:latest -f Dockerfile --push .
-	docker buildx build --platform linux/arm64,linux/amd64 -t fosrl/gerbil:$(tag) -f Dockerfile --push .
+	docker buildx build --platform linux/arm64,linux/amd64 -t fosrl/gerbil:$(tag) -f Dockerfile --build-arg GO_VERSION=$(GO_VERSION) --push .
 
 build:
-	docker build -t fosrl/gerbil:latest .
+	docker build -t fosrl/gerbil:latest --build-arg GO_VERSION=$(GO_VERSION) .
 
 push:
 	docker push fosrl/gerbil:latest

README.md

@@ -40,6 +40,24 @@ The PROXY protocol allows downstream proxies to know the real client IP address
 
 In single node (self hosted) Pangolin deployments this can be bypassed by using port 443:443 to route to Traefik instead of the SNI proxy at 8443.
 
+### Observability with OpenTelemetry
+
+Gerbil includes comprehensive OpenTelemetry metrics instrumentation for monitoring and observability. Metrics can be exported via:
+
+- **Prometheus**: Pull-based metrics at the `/metrics` endpoint (enabled by default)
+- **OTLP**: Push-based metrics to any OpenTelemetry-compatible collector
+
+Key metrics include:
+
+- WireGuard interface and peer status
+- Bandwidth usage per peer
+- Active relay sessions and proxy connections
+- Handshake success/failure rates
+- Route lookup cache hit/miss ratios
+- Go runtime metrics (GC, goroutines, memory)
+
+See [docs/observability.md](docs/observability.md) for complete documentation, metrics reference, and examples.
+
 ## CLI Args
 
 Important:
@@ -120,7 +138,7 @@ make
 
 ### Binary
 
-Make sure to have Go 1.23.1 installed.
+Make sure to have Go 1.26 installed.
 
 ```bash
 make local

docs/observability.md

@@ -0,0 +1,273 @@
+<!-- markdownlint-disable MD036 MD060 -->
+# Gerbil Observability Architecture
+
+This document describes the metrics subsystem for Gerbil, explains the design
+decisions, and shows how to configure each backend.
+
+---
+
+## Architecture Overview
+
+Gerbil's metrics subsystem uses a **pluggable backend** design:
+
+```text
+main.go  ─── internal/metrics  ─── internal/observability  ─── backend
+                 (facade)                 (interface)           Prometheus
+                                                           OR   OTel/OTLP
+                                                           OR   Noop (disabled)
+```
+
+Application code (main, relay, proxy) calls only the `metrics.Record*`
+functions in `internal/metrics`. That package delegates to whichever backend
+was selected at startup via `internal/observability.Backend`.
+
+### Why Prometheus-native and OTel are mutually exclusive
+
+**Exactly one** metrics backend may be active at runtime:
+
+| Mode | What happens |
+|------|-------------|
+| `prometheus` | Native Prometheus client registers metrics on a dedicated registry and exposes `/metrics`. No OTel SDK is initialised. |
+| `otel` | OTel SDK pushes metrics via OTLP/gRPC or OTLP/HTTP to an external collector. No `/metrics` endpoint is exposed. |
+| `none` | A safe noop backend is used. All `Record*` calls are discarded. |
+
+Running both simultaneously would mean every metric is recorded twice through
+two different code paths, with differing semantics (pull vs. push, different
+naming rules, different cardinality handling). The design enforces a single
+source of truth.
+
+### Future OTel tracing and logging
+
+The `internal/observability/otel/` package is designed so that tracing and
+logging support can be added **beside** the existing metrics code without
+touching the Prometheus-native path:
+
+```bash
+internal/observability/otel/
+  backend.go     ← metrics
+  exporter.go    ← OTLP exporter creation
+  resource.go    ← OTel resource
+  trace.go       ← future: TracerProvider setup
+  log.go         ← future: LoggerProvider setup
+```
+
+---
+
+## Configuration
+
+### Config precedence
+
+1. CLI flags (highest priority)
+2. Environment variables
+3. Defaults
+
+### Config struct
+
+```go
+type MetricsConfig struct {
+    Enabled               bool
+    Backend               string // "prometheus" | "otel" | "none"
+    Prometheus            PrometheusConfig
+    OTel                  OTelConfig
+    ServiceName           string
+    ServiceVersion        string
+    DeploymentEnvironment string
+}
+
+type PrometheusConfig struct {
+    Path string // default: "/metrics"
+}
+
+type OTelConfig struct {
+    Protocol       string        // "grpc" (default) or "http"
+    Endpoint       string        // default: "localhost:4317"
+    Insecure       bool          // default: true
+    ExportInterval time.Duration // default: 60s
+    Timeout        time.Duration // default: 10s
+}
+```
+
+### Environment variables
+
+| Variable | Default | Description |
+|----------|---------|-------------|
+| `METRICS_ENABLED` | `true` | Enable/disable metrics |
+| `METRICS_BACKEND` | `prometheus` | Backend: `prometheus`, `otel`, or `none` |
+| `METRICS_PATH` | `/metrics` | HTTP path for Prometheus endpoint |
+| `OTEL_METRICS_PROTOCOL` | `grpc` | OTLP transport: `grpc` or `http` |
+| `OTEL_METRICS_ENDPOINT` | `localhost:4317` | OTLP collector address |
+| `OTEL_METRICS_INSECURE` | `true` | Disable TLS for OTLP |
+| `OTEL_METRICS_EXPORT_INTERVAL` | `60s` | Push interval (e.g. `10s`, `1m`) |
+| `OTEL_METRICS_TIMEOUT` | `10s` | Timeout for OTLP exporter connection setup |
+| `DEPLOYMENT_ENVIRONMENT` | _(unset)_ | OTel deployment.environment attribute |
+
+### CLI flags
+
+```bash
+--metrics-enabled            bool    (default: true)
+--metrics-backend            string  (default: prometheus)
+--metrics-path               string  (default: /metrics)
+--otel-metrics-protocol      string  (default: grpc)
+--otel-metrics-endpoint      string  (default: localhost:4317)
+--otel-metrics-insecure      bool    (default: true)
+--otel-metrics-export-interval  duration  (default: 60s)
+--otel-metrics-timeout          duration  (default: 10s)
+```
+
+---
+
+## When to choose each backend
+
+| Criterion | Prometheus | OTel/OTLP |
+|-----------|-----------|-----------|
+| Existing Prometheus/Grafana stack | ✅ | |
+| Pull-based scraping | ✅ | |
+| No external collector required | ✅ | |
+| Vendor-neutral telemetry | | ✅ |
+| Push-based export | | ✅ |
+| Grafana Cloud / managed OTLP | | ✅ |
+| Future traces + logs via same pipeline | | ✅ |
+
+---
+
+## Enabling Prometheus-native mode
+
+### Environment variables
+
+```bash
+METRICS_ENABLED=true
+METRICS_BACKEND=prometheus
+METRICS_PATH=/metrics
+```
+
+### CLI
+
+```bash
+./gerbil --metrics-enabled --metrics-backend=prometheus --metrics-path=/metrics \
+         --config=/etc/gerbil/config.json
+```
+
+The metrics config is supplied separately via env/flags; it is not embedded
+in the WireGuard config file.
+
+The Prometheus `/metrics` endpoint is registered only when
+`--metrics-backend=prometheus`. All gerbil_* metrics plus Go runtime metrics
+are available.
+
+---
+
+## Enabling OTel mode
+
+### Environment variables
+
+```bash
+export METRICS_ENABLED=true
+export METRICS_BACKEND=otel
+export OTEL_METRICS_PROTOCOL=grpc
+export OTEL_METRICS_ENDPOINT=otel-collector:4317
+export OTEL_METRICS_INSECURE=true
+export OTEL_METRICS_EXPORT_INTERVAL=10s
+export OTEL_METRICS_TIMEOUT=10s
+export DEPLOYMENT_ENVIRONMENT=production
+```
+
+### CLI
+
+```bash
+./gerbil --metrics-enabled \
+         --metrics-backend=otel \
+         --otel-metrics-protocol=grpc \
+         --otel-metrics-endpoint=otel-collector:4317 \
+         --otel-metrics-insecure \
+         --otel-metrics-export-interval=10s \
+         --otel-metrics-timeout=10s \
+         --config=/etc/gerbil/config.json
+```
+
+### HTTP mode (OTLP/HTTP)
+
+```bash
+export OTEL_METRICS_PROTOCOL=http
+export OTEL_METRICS_ENDPOINT=otel-collector:4318
+```
+
+---
+
+## Disabling metrics
+
+```bash
+export METRICS_ENABLED=false
+# or
+./gerbil --metrics-enabled=false
+# or
+./gerbil --metrics-backend=none
+```
+
+When disabled, all `Record*` calls are directed to a safe noop backend that
+discards observations without allocating or locking.
+
+---
+
+## Metric catalog
+
+All metrics use the prefix `gerbil_<component>_<name>`.
+
+### WireGuard metrics
+
+| Metric | Type | Labels | Description |
+|--------|------|--------|-------------|
+| `gerbil_wg_interface_up` | Gauge | `ifname`, `instance` | 1=up, 0=down |
+| `gerbil_wg_peers_total` | UpDownCounter | `ifname` | Configured peers |
+| `gerbil_wg_peer_connected` | Gauge | `ifname`, `peer` | 1=connected, 0=disconnected |
+| `gerbil_wg_bytes_received_total` | Counter | `ifname`, `peer` | Bytes received |
+| `gerbil_wg_bytes_transmitted_total` | Counter | `ifname`, `peer` | Bytes transmitted |
+| `gerbil_wg_handshakes_total` | Counter | `ifname`, `peer`, `result` | Handshake attempts |
+| `gerbil_wg_handshake_latency_seconds` | Histogram | `ifname`, `peer` | Handshake duration |
+| `gerbil_wg_peer_rtt_seconds` | Histogram | `ifname`, `peer` | Peer round-trip time |
+
+### Relay metrics
+
+| Metric | Type | Labels |
+|--------|------|--------|
+| `gerbil_proxy_mapping_active` | UpDownCounter | `ifname` |
+| `gerbil_active_sessions` | UpDownCounter | `ifname` |
+| `gerbil_udp_packets_total` | Counter | `ifname`, `type`, `direction` |
+| `gerbil_hole_punch_events_total` | Counter | `ifname`, `result` |
+
+### SNI proxy metrics
+
+| Metric | Type | Labels |
+|--------|------|--------|
+| `gerbil_sni_connections_total` | Counter | `result` |
+| `gerbil_sni_active_connections` | UpDownCounter | _(none)_ |
+| `gerbil_sni_route_cache_hits_total` | Counter | `result` |
+| `gerbil_sni_route_api_requests_total` | Counter | `result` |
+| `gerbil_proxy_route_lookups_total` | Counter | `result`, `hostname` |
+
+### HTTP metrics
+
+| Metric | Type | Labels |
+|--------|------|--------|
+| `gerbil_http_requests_total` | Counter | `endpoint`, `method`, `status_code` |
+| `gerbil_http_request_duration_seconds` | Histogram | `endpoint`, `method` |
+
+---
+
+## Using Docker Compose
+
+The `docker-compose.metrics.yml` provides a complete observability stack.
+
+**Prometheus mode:**
+
+```bash
+METRICS_BACKEND=prometheus docker-compose -f docker compose.metrics.yml up -d
+# Scrape at http://localhost:3003/metrics
+# Grafana at http://localhost:3000 (admin/admin)
+```
+
+**OTel mode:**
+
+```bash
+METRICS_BACKEND=otel OTEL_METRICS_ENDPOINT=otel-collector:4317 \
+  docker compose -f docker-compose.metrics.yml up -d
+```

examples/otel-collector-config.yaml

@@ -0,0 +1,47 @@
+file_format: '1.0'
+receivers:
+  otlp:
+    protocols:
+      grpc:
+        endpoint: 0.0.0.0:4317
+      http:
+        endpoint: 0.0.0.0:4318
+
+processors:
+  batch:
+    timeout: 10s
+    send_batch_size: 1024
+  
+  # Add resource attributes
+  resource:
+    attributes:
+      - key: service.environment
+        value: "development"
+        action: insert
+
+exporters:
+  # Prometheus exporter for scraping
+  prometheus:
+    endpoint: "0.0.0.0:8889"
+    namespace: "gerbil"
+    send_timestamps: true
+    metric_expiration: 5m
+    resource_to_telemetry_conversion:
+      enabled: true
+
+  # Prometheus remote write (optional)
+  prometheusremotewrite:
+    endpoint: "http://prometheus:9090/api/v1/write"
+    tls:
+      insecure: true
+  
+  # Debug exporter for debugging
+  debug:
+    verbosity: normal
+
+service:
+  pipelines:
+    metrics:
+      receivers: [otlp]
+      processors: [batch, resource]
+      exporters: [prometheus, prometheusremotewrite, debug]

examples/prometheus.yml

@@ -0,0 +1,24 @@
+global:
+  scrape_interval: 15s
+  evaluation_interval: 15s
+  external_labels:
+    cluster: 'gerbil-dev'
+
+scrape_configs:
+  # Scrape Gerbil's /metrics endpoint directly
+  - job_name: 'gerbil'
+    static_configs:
+      - targets: ['gerbil:3003']
+        labels:
+          service: 'gerbil'
+          environment: 'development'
+
+  # Scrape OpenTelemetry Collector metrics
+  - job_name: 'otel-collector'
+    static_configs:
+      - targets: ['otel-collector:8888']
+        labels:
+          service: 'otel-collector'
+      - targets: ['otel-collector:8889']
+        labels:
+          service: 'otel-collector-prometheus-exporter'

go.mod

@@ -1,23 +1,49 @@
 module github.com/fosrl/gerbil
 
-go 1.25
+go 1.26.0
 
 require (
 	github.com/patrickmn/go-cache v2.1.0+incompatible
+	github.com/prometheus/client_golang v1.24.1
 	github.com/vishvananda/netlink v1.3.1
-	golang.org/x/crypto v0.46.0
-	golang.org/x/sync v0.1.0
+	go.opentelemetry.io/otel v1.46.0
+	go.opentelemetry.io/otel/exporters/otlp/otlpmetric/otlpmetricgrpc v1.46.0
+	go.opentelemetry.io/otel/exporters/otlp/otlpmetric/otlpmetrichttp v1.46.0
+	go.opentelemetry.io/otel/metric v1.46.0
+	go.opentelemetry.io/otel/sdk v1.46.0
+	go.opentelemetry.io/otel/sdk/metric v1.46.0
+	golang.org/x/crypto v0.55.0
+	golang.org/x/sync v0.22.0
 	golang.zx2c4.com/wireguard/wgctrl v0.0.0-20230429144221-925a1e7659e6
 )
 
 require (
-	github.com/google/go-cmp v0.5.9 // indirect
+	github.com/beorn7/perks v1.0.1 // indirect
+	github.com/cenkalti/backoff/v5 v5.0.3 // indirect
+	github.com/cespare/xxhash/v2 v2.3.0 // indirect
+	github.com/go-logr/logr v1.4.4 // indirect
+	github.com/go-logr/stdr v1.2.2 // indirect
+	github.com/google/go-cmp v0.7.0 // indirect
+	github.com/google/uuid v1.6.0 // indirect
+	github.com/grpc-ecosystem/grpc-gateway/v2 v2.30.0 // indirect
 	github.com/josharian/native v1.1.0 // indirect
 	github.com/mdlayher/genetlink v1.3.2 // indirect
 	github.com/mdlayher/netlink v1.7.2 // indirect
 	github.com/mdlayher/socket v0.4.1 // indirect
+	github.com/munnerz/goautoneg v0.0.0-20191010083416-a7dc8b61c822 // indirect
+	github.com/prometheus/client_model v0.6.2 // indirect
+	github.com/prometheus/common v0.70.1 // indirect
+	github.com/prometheus/procfs v0.21.1 // indirect
 	github.com/vishvananda/netns v0.0.5 // indirect
-	golang.org/x/net v0.47.0 // indirect
-	golang.org/x/sys v0.39.0 // indirect
+	go.opentelemetry.io/auto/sdk v1.2.1 // indirect
+	go.opentelemetry.io/otel/trace v1.46.0 // indirect
+	go.opentelemetry.io/proto/otlp v1.11.0 // indirect
+	golang.org/x/net v0.58.0 // indirect
+	golang.org/x/sys v0.47.0 // indirect
+	golang.org/x/text v0.41.0 // indirect
 	golang.zx2c4.com/wireguard v0.0.0-20230325221338-052af4a8072b // indirect
+	google.golang.org/genproto/googleapis/api v0.0.0-20260819154853-08b0e4226688 // indirect
+	google.golang.org/genproto/googleapis/rpc v0.0.0-20260819154853-08b0e4226688 // indirect
+	google.golang.org/grpc v1.83.1 // indirect
+	google.golang.org/protobuf v1.36.12 // indirect
 )

go.sum

@@ -1,7 +1,28 @@
-github.com/google/go-cmp v0.5.9 h1:O2Tfq5qg4qc4AmwVlvv0oLiVAGB7enBSJ2x2DqQFi38=
-github.com/google/go-cmp v0.5.9/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=
+github.com/beorn7/perks v1.0.1 h1:VlbKKnNfV8bJzeqoa4cOKqO6bYr3WgKZxO8Z16+hsOM=
+github.com/beorn7/perks v1.0.1/go.mod h1:G2ZrVWU2WbWT9wwq4/hrbKbnv/1ERSJQ0ibhJ6rlkpw=
+github.com/cenkalti/backoff/v5 v5.0.3 h1:ZN+IMa753KfX5hd8vVaMixjnqRZ3y8CuJKRKj1xcsSM=
+github.com/cenkalti/backoff/v5 v5.0.3/go.mod h1:rkhZdG3JZukswDf7f0cwqPNk4K0sa+F97BxZthm/crw=
+github.com/cespare/xxhash/v2 v2.3.0 h1:UL815xU9SqsFlibzuggzjXhog7bL6oX9BbNZnL2UFvs=
+github.com/cespare/xxhash/v2 v2.3.0/go.mod h1:VGX0DQ3Q6kWi7AoAeZDth3/j3BFtOZR5XLFGgcrjCOs=
+github.com/go-logr/logr v1.2.2/go.mod h1:jdQByPbusPIv2/zmleS9BjJVeZ6kBagPoEUsqbVz/1A=
+github.com/go-logr/logr v1.4.4 h1:tG4xh9yMsRCAiodLVTxyrkzSZ9+o0L1Kg/+cPVcbP/8=
+github.com/go-logr/logr v1.4.4/go.mod h1:9T104GzyrTigFIr8wt5mBrctHMim0Nb2HLGrmQ40KvY=
+github.com/go-logr/stdr v1.2.2 h1:hSWxHoqTgW2S2qGc0LTAI563KZ5YKYRhT3MFKZMbjag=
+github.com/go-logr/stdr v1.2.2/go.mod h1:mMo/vtBO5dYbehREoey6XUKy/eSumjCCveDpRre4VKE=
+github.com/golang/protobuf v1.5.4 h1:i7eJL8qZTpSEXOPTxNKhASYpMn+8e5Q6AdndVa1dWek=
+github.com/golang/protobuf v1.5.4/go.mod h1:lnTiLA8Wa4RWRcIUkrtSVa5nRhsEGBg48fD6rSs7xps=
+github.com/google/go-cmp v0.7.0 h1:wk8382ETsv4JYUZwIsn6YpYiWiBsYLSJiTsyBybVuN8=
+github.com/google/go-cmp v0.7.0/go.mod h1:pXiqmnSA92OHEEa9HXL2W4E7lf9JzCmGVUdgjX3N/iU=
+github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=
+github.com/google/uuid v1.6.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
+github.com/grpc-ecosystem/grpc-gateway/v2 v2.30.0 h1:/Tnpcb2E0Pz/tN9s3bfEY2Q8ePCEX9iuS+cneUwncnw=
+github.com/grpc-ecosystem/grpc-gateway/v2 v2.30.0/go.mod h1:zOBXOsUaBSjKgmH4OGzV1esUpR3oUSCPYVd2cUBjKYY=
 github.com/josharian/native v1.1.0 h1:uuaP0hAbW7Y4l0ZRQ6C9zfb7Mg1mbFKry/xzDAfmtLA=
 github.com/josharian/native v1.1.0/go.mod h1:7X/raswPFr05uY3HiLlYeyQntB6OO7E/d2Cu7qoaN2w=
+github.com/klauspost/compress v1.19.1 h1:VsB4HPswih7mmZ8WleSFQ75c/Ui1M4trX5oAsJnhSlk=
+github.com/klauspost/compress v1.19.1/go.mod h1:cwPg85FWrGar70rWktvGQj8/hthj3wpl0PGDogxkrSQ=
+github.com/kylelemons/godebug v1.1.0 h1:RPNrshWIDI6G2gRW9EHilWtl7Z6Sb1BR0xunSBf0SNc=
+github.com/kylelemons/godebug v1.1.0/go.mod h1:9/0rRGxNHcop5bhtWyNeEfOS8JIWk580+fNqagV/RAw=
 github.com/mdlayher/genetlink v1.3.2 h1:KdrNKe+CTu+IbZnm/GVUMXSqBBLqcGpRDa0xkQy56gw=
 github.com/mdlayher/genetlink v1.3.2/go.mod h1:tcC3pkCrPUGIKKsCsp0B3AdaaKuHtaxoJRz3cc+528o=
 github.com/mdlayher/netlink v1.7.2 h1:/UtM3ofJap7Vl4QWCPDGXY8d3GIY2UGSDbK+QWmY8/g=
@@ -10,23 +31,73 @@ github.com/mdlayher/socket v0.4.1 h1:eM9y2/jlbs1M615oshPQOHZzj6R6wMT7bX5NPiQvn2U
 github.com/mdlayher/socket v0.4.1/go.mod h1:cAqeGjoufqdxWkD7DkpyS+wcefOtmu5OQ8KuoJGIReA=
 github.com/mikioh/ipaddr v0.0.0-20190404000644-d465c8ab6721 h1:RlZweED6sbSArvlE924+mUcZuXKLBHA35U7LN621Bws=
 github.com/mikioh/ipaddr v0.0.0-20190404000644-d465c8ab6721/go.mod h1:Ickgr2WtCLZ2MDGd4Gr0geeCH5HybhRJbonOgQpvSxc=
+github.com/munnerz/goautoneg v0.0.0-20191010083416-a7dc8b61c822 h1:C3w9PqII01/Oq1c1nUAm88MOHcQC9l5mIlSMApZMrHA=
+github.com/munnerz/goautoneg v0.0.0-20191010083416-a7dc8b61c822/go.mod h1:+n7T8mK8HuQTcFwEeznm/DIxMOiR9yIdICNftLE1DvQ=
 github.com/patrickmn/go-cache v2.1.0+incompatible h1:HRMgzkcYKYpi3C8ajMPV8OFXaaRUnok+kx1WdO15EQc=
 github.com/patrickmn/go-cache v2.1.0+incompatible/go.mod h1:3Qf8kWWT7OJRJbdiICTKqZju1ZixQ/KpMGzzAfe6+WQ=
+github.com/prometheus/client_golang v1.24.1 h1:JnJkREXzWxUdCuPFpIWZiPispT9xVV59uiuyR2bPlnU=
+github.com/prometheus/client_golang v1.24.1/go.mod h1:F+oSRECHg4sse5ucfYpYDeIv/hu68Zo0uoHKetWnzcE=
+github.com/prometheus/client_model v0.6.2 h1:oBsgwpGs7iVziMvrGhE53c/GrLUsZdHnqNwqPLxwZyk=
+github.com/prometheus/client_model v0.6.2/go.mod h1:y3m2F6Gdpfy6Ut/GBsUqTWZqCUvMVzSfMLjcu6wAwpE=
+github.com/prometheus/common v0.70.1 h1:1HvjP4D5oL3t8RsPlwxA9onvvStjtIHYE5XuuwOi/PY=
+github.com/prometheus/common v0.70.1/go.mod h1:VdFUQDMZK3VLkurFUVhia6uys/0suUp86TJz5qbJRhc=
+github.com/prometheus/procfs v0.21.1 h1:GljZCt+zSTS+NZq88cyQ1LjZ+RCHp3uVuabBWA5+OJI=
+github.com/prometheus/procfs v0.21.1/go.mod h1:aB55Cww9pdSJVHk0hUf0inxWyyjPogFIjmHKYgMKmtY=
+github.com/stretchr/testify v1.12.1 h1:EuwCh5fleGS7H32xRwO3wRGT7DxrDhLAT6FF8MpWDWE=
+github.com/stretchr/testify v1.12.1/go.mod h1:MDEgiDPPsNp5cuIrHPPCyornHKgEVbtFUmoNlxoYthg=
 github.com/vishvananda/netlink v1.3.1 h1:3AEMt62VKqz90r0tmNhog0r/PpWKmrEShJU0wJW6bV0=
 github.com/vishvananda/netlink v1.3.1/go.mod h1:ARtKouGSTGchR8aMwmkzC0qiNPrrWO5JS/XMVl45+b4=
 github.com/vishvananda/netns v0.0.5 h1:DfiHV+j8bA32MFM7bfEunvT8IAqQ/NzSJHtcmW5zdEY=
 github.com/vishvananda/netns v0.0.5/go.mod h1:SpkAiCQRtJ6TvvxPnOSyH3BMl6unz3xZlaprSwhNNJM=
-golang.org/x/crypto v0.46.0 h1:cKRW/pmt1pKAfetfu+RCEvjvZkA9RimPbh7bhFjGVBU=
-golang.org/x/crypto v0.46.0/go.mod h1:Evb/oLKmMraqjZ2iQTwDwvCtJkczlDuTmdJXoZVzqU0=
-golang.org/x/net v0.47.0 h1:Mx+4dIFzqraBXUugkia1OOvlD6LemFo1ALMHjrXDOhY=
-golang.org/x/net v0.47.0/go.mod h1:/jNxtkgq5yWUGYkaZGqo27cfGZ1c5Nen03aYrrKpVRU=
-golang.org/x/sync v0.1.0 h1:wsuoTGHzEhffawBOhz5CYhcrV4IdKZbEyZjBMuTp12o=
-golang.org/x/sync v0.1.0/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
+go.opentelemetry.io/auto/sdk v1.2.1 h1:jXsnJ4Lmnqd11kwkBV2LgLoFMZKizbCi5fNZ/ipaZ64=
+go.opentelemetry.io/auto/sdk v1.2.1/go.mod h1:KRTj+aOaElaLi+wW1kO/DZRXwkF4C5xPbEe3ZiIhN7Y=
+go.opentelemetry.io/otel v1.46.0 h1:FHt5/CDyVxi/8IM1CH7VE/rRgq3kLHa2mSTVMO8AWyc=
+go.opentelemetry.io/otel v1.46.0/go.mod h1:Gj3SEScelsNC45tp4nSxRYlS+f5iez7W8XPMCt905kE=
+go.opentelemetry.io/otel/exporters/otlp/otlpmetric/otlpmetricgrpc v1.46.0 h1:qkDYCAFiZXLcs1L4aY+tP2wguQ4kURANqHOQMA2et2s=
+go.opentelemetry.io/otel/exporters/otlp/otlpmetric/otlpmetricgrpc v1.46.0/go.mod h1:tkipS4DRzmpAmvg+Gw4++O1IdDq6TVDnvnYU6cmbQVs=
+go.opentelemetry.io/otel/exporters/otlp/otlpmetric/otlpmetrichttp v1.46.0 h1:AP23h/mFgb/lc7tdck1Kfn9qxsM8TAeNPCU5C3pzaps=
+go.opentelemetry.io/otel/exporters/otlp/otlpmetric/otlpmetrichttp v1.46.0/go.mod h1:K4EqCe1b4kGk5WR690ntg9LaBfsPoV32FwthbyoptuA=
+go.opentelemetry.io/otel/metric v1.46.0 h1:yBnkXvgV7AXFILZc5K6IZe/CBFF3OS7BJ8ov6/lj0K8=
+go.opentelemetry.io/otel/metric v1.46.0/go.mod h1:iPmdWqifKUdzziPkvvzIJXITl56fQx2mGM/DHLB3/2o=
+go.opentelemetry.io/otel/metric/x v0.68.0 h1:TA/cBT23D3MnxYPwHL7YFOdYGdx0A0v+s7Mzotpd1dU=
+go.opentelemetry.io/otel/metric/x v0.68.0/go.mod h1:agudOmvWhwUTjgibWDzxD2PoWYnpw5Ht5jISYOD2Hd4=
+go.opentelemetry.io/otel/sdk v1.46.0 h1:h5CNQQjEbuQXY/JfZtgt3i7HVFV3aHPO2OAwO2eTYPI=
+go.opentelemetry.io/otel/sdk v1.46.0/go.mod h1:GAERFXFt5SYCEB+YiKUbMBeza6UaDH7GmGOZEfh2gSM=
+go.opentelemetry.io/otel/sdk/metric v1.46.0 h1:0piZ26EG4RBfebb2jhDH6ERCYHoVWduc3kLgPCwSnSE=
+go.opentelemetry.io/otel/sdk/metric v1.46.0/go.mod h1:I1PbKrdVc8Qu8HYVDNtqVIwLwjNrhsV/uFuxfwg8mO4=
+go.opentelemetry.io/otel/trace v1.46.0 h1:OULy7ccdJnZtJ0UDYFOIGaCmiWzJ8Vi2G/Rsu60qs1c=
+go.opentelemetry.io/otel/trace v1.46.0/go.mod h1:J7GAXweO77XSFkB/rmAqk9D6ihszhFjLU+d9WuUxDLI=
+go.opentelemetry.io/proto/otlp v1.11.0 h1:5rrYs0Ykyj50sdU/JU0x8etU+LubXWb+gED6TbEdMIk=
+go.opentelemetry.io/proto/otlp v1.11.0/go.mod h1:SmVizdCOAm3XBtG1g1NnOdhW6jtddT72hLMhv8VwA8E=
+go.uber.org/goleak v1.3.0 h1:2K3zAYmnTNqV73imy9J1T3WC+gmCePx2hEGkimedGto=
+go.uber.org/goleak v1.3.0/go.mod h1:CoHD4mav9JJNrW/WLlf7HGZPjdw8EucARQHekz1X6bE=
+go.yaml.in/yaml/v2 v2.4.4 h1:tuyd0P+2Ont/d6e2rl3be67goVK4R6deVxCUX5vyPaQ=
+go.yaml.in/yaml/v2 v2.4.4/go.mod h1:gMZqIpDtDqOfM0uNfy0SkpRhvUryYH0Z6wdMYcacYXQ=
+go.yaml.in/yaml/v3 v3.0.5 h1:N6y/pJk8buWs9NY5ERU2HSMfm+IuD/OtfdAnq6kESPw=
+go.yaml.in/yaml/v3 v3.0.5/go.mod h1:HVTZu1O7/Vkt2N+BFy8Zza+lnLsABggaTM2ZpNIGuKg=
+golang.org/x/crypto v0.55.0 h1:+KWHjbgOaAQ66dh/YlkZKHlz9ZUlq61AFirAR9ntP8M=
+golang.org/x/crypto v0.55.0/go.mod h1:uq0V9dE/fzQuJtbnL+2EhWOE63vo164FY8xqEnV9xis=
+golang.org/x/net v0.58.0 h1:ynWG7rqYi4ccpTEuPZ2QGWHktVEM9DMCj9yzDE0Q7To=
+golang.org/x/net v0.58.0/go.mod h1:YwCddHnFlT7eLQqVprV19OnhLGtc5xOKgE0RyqgfWAU=
+golang.org/x/sync v0.22.0 h1:SZjpbeLmrCk4xhRSZFNZW5gFUeCeFgjekvI/+gfScek=
+golang.org/x/sync v0.22.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
 golang.org/x/sys v0.2.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.10.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.39.0 h1:CvCKL8MeisomCi6qNZ+wbb0DN9E5AATixKsvNtMoMFk=
-golang.org/x/sys v0.39.0/go.mod h1:OgkHotnGiDImocRcuBABYBEXf8A9a87e/uXjp9XT3ks=
+golang.org/x/sys v0.47.0 h1:o7XGOvZQCADBQQ4Y7VNq2dRWQR7JmOUW8Kxx4ZsNgWs=
+golang.org/x/sys v0.47.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
+golang.org/x/text v0.41.0 h1:vz/seA0lnX87Othu2f/0L24RcgrXD9/YFTSuGjj3rH8=
+golang.org/x/text v0.41.0/go.mod h1:jvf1O8ajNzZqhSrQBPbutR/EB83Cc0CFrezNQIwbb5M=
 golang.zx2c4.com/wireguard v0.0.0-20230325221338-052af4a8072b h1:J1CaxgLerRR5lgx3wnr6L04cJFbWoceSK9JWBdglINo=
 golang.zx2c4.com/wireguard v0.0.0-20230325221338-052af4a8072b/go.mod h1:tqur9LnfstdR9ep2LaJT4lFUl0EjlHtge+gAjmsHUG4=
 golang.zx2c4.com/wireguard/wgctrl v0.0.0-20230429144221-925a1e7659e6 h1:CawjfCvYQH2OU3/TnxLx97WDSUDRABfT18pCOYwc2GE=
 golang.zx2c4.com/wireguard/wgctrl v0.0.0-20230429144221-925a1e7659e6/go.mod h1:3rxYc4HtVcSG9gVaTs2GEBdehh+sYPOwKtyUWEOTb80=
+gonum.org/v1/gonum v0.17.0 h1:VbpOemQlsSMrYmn7T2OUvQ4dqxQXU+ouZFQsZOx50z4=
+gonum.org/v1/gonum v0.17.0/go.mod h1:El3tOrEuMpv2UdMrbNlKEh9vd86bmQ6vqIcDwxEOc1E=
+google.golang.org/genproto/googleapis/api v0.0.0-20260819154853-08b0e4226688 h1:ax2KzoSRIZU/M0cIxri3pKxy99vniH1PVxWC6si/eZI=
+google.golang.org/genproto/googleapis/api v0.0.0-20260819154853-08b0e4226688/go.mod h1:1RJ9BQGyNdZwkGc1eTqkErfRZ6RJyYPHZo73BZ1vQqI=
+google.golang.org/genproto/googleapis/rpc v0.0.0-20260819154853-08b0e4226688 h1:cYNAzI2sUwhmCcoj9TxvihSrqsxt6uIkj3rDRhSDmW4=
+google.golang.org/genproto/googleapis/rpc v0.0.0-20260819154853-08b0e4226688/go.mod h1:DjtHYE8FKJLivXcBEjGwndXfIC23G0VpXiXKqG179uA=
+google.golang.org/grpc v1.83.1 h1:HIO0+BEtBP6soyqvqC8sNUjZ7bTs+0hFQuFF+RAy++Y=
+google.golang.org/grpc v1.83.1/go.mod h1:kDyl6SKsiHKt0uylY5gtn5cEjkrIOhQOGDgIc4JGwzQ=
+google.golang.org/protobuf v1.36.12 h1:pJOKDDOyeXErUroCihFAd5LQuwXBSpVnKGrj5o/fwxc=
+google.golang.org/protobuf v1.36.12/go.mod h1:HTf+CrKn2C3g5S8VImy6tdcUvCska2kB7j23XfzDpco=

internal/metrics/metrics.go

@@ -0,0 +1,921 @@
+// Package metrics provides the application-level metrics facade for Gerbil.
+//
+// Application code (main, relay, proxy) uses only the Record* functions in this
+// package. The actual recording is delegated to the backend selected in
+// internal/observability. Neither Prometheus nor OTel packages are imported here.
+package metrics
+
+import (
+	"context"
+	"fmt"
+	"net/http"
+	"sync"
+
+	"github.com/fosrl/gerbil/internal/observability"
+)
+
+// Config is the metrics configuration type. It is an alias for
+// observability.MetricsConfig so callers do not need to import observability.
+type Config = observability.MetricsConfig
+
+// PrometheusConfig is re-exported for convenience.
+type PrometheusConfig = observability.PrometheusConfig
+
+// OTelConfig is re-exported for convenience.
+type OTelConfig = observability.OTelConfig
+
+var (
+	backend observability.Backend
+	initMu  sync.Mutex
+
+	// Interface and peer metrics
+	wgInterfaceUp      observability.Int64Gauge
+	wgPeersTotal       observability.UpDownCounter
+	wgPeerConnected    observability.Int64Gauge
+	wgHandshakesTotal  observability.Counter
+	wgHandshakeLatency observability.Histogram
+	wgPeerRTT          observability.Histogram
+	wgBytesReceived    observability.Counter
+	wgBytesTransmitted observability.Counter
+	allowedIPsCount    observability.UpDownCounter
+	keyRotationTotal   observability.Counter
+
+	// System and proxy metrics
+	netlinkEventsTotal     observability.Counter
+	netlinkErrorsTotal     observability.Counter
+	syncDuration           observability.Histogram
+	workqueueDepth         observability.UpDownCounter
+	kernelModuleLoads      observability.Counter
+	firewallRulesApplied   observability.Counter
+	activeSessions         observability.UpDownCounter
+	activeProxyConnections observability.UpDownCounter
+	proxyRouteLookups      observability.Counter
+	proxyTLSHandshake      observability.Histogram
+	proxyBytesTransmitted  observability.Counter
+
+	// UDP Relay / Proxy Metrics
+	udpPacketsTotal            observability.Counter
+	udpPacketSizeBytes         observability.Histogram
+	holePunchEventsTotal       observability.Counter
+	proxyMappingActive         observability.UpDownCounter
+	relayUDPConnectionsActive  observability.UpDownCounter
+	sessionRebuiltTotal        observability.Counter
+	commPatternActive          observability.UpDownCounter
+	proxyCleanupRemovedTotal   observability.Counter
+	proxyConnectionErrorsTotal observability.Counter
+	proxyInitialMappingsTotal  observability.Int64Gauge
+	proxyMappingUpdatesTotal   observability.Counter
+	proxyIdleCleanupDuration   observability.Histogram
+
+	// SNI Proxy Metrics
+	sniConnectionsTotal              observability.Counter
+	sniConnectionDuration            observability.Histogram
+	sniActiveConnections             observability.UpDownCounter
+	sniRouteCacheHitsTotal           observability.Counter
+	sniRouteAPIRequestsTotal         observability.Counter
+	sniRouteAPILatency               observability.Histogram
+	sniLocalOverrideTotal            observability.Counter
+	sniTrustedProxyEventsTotal       observability.Counter
+	sniProxyProtocolParseErrorsTotal observability.Counter
+	sniDataBytesTotal                observability.Counter
+	sniTunnelTerminationsTotal       observability.Counter
+
+	// HTTP API & Peer Management Metrics
+	httpRequestsTotal               observability.Counter
+	httpRequestDuration             observability.Histogram
+	peerOperationsTotal             observability.Counter
+	proxyMappingUpdateRequestsTotal observability.Counter
+	destinationsUpdateRequestsTotal observability.Counter
+
+	// Remote Configuration, Reporting & Housekeeping
+	remoteConfigFetchesTotal observability.Counter
+	bandwidthReportsTotal    observability.Counter
+	peerBandwidthBytesTotal  observability.Counter
+	memorySpikeTotal         observability.Counter
+	heapProfilesWrittenTotal observability.Counter
+
+	// Operational metrics
+	configReloadsTotal    observability.Counter
+	restartTotal          observability.Counter
+	authFailuresTotal     observability.Counter
+	aclDeniedTotal        observability.Counter
+	certificateExpiryDays observability.Float64Gauge
+)
+
+// DefaultConfig returns a default metrics configuration.
+func DefaultConfig() Config {
+	return observability.DefaultMetricsConfig()
+}
+
+// Initialize sets up the metrics system using the selected backend.
+// It returns the /metrics HTTP handler (non-nil only for Prometheus backend).
+func Initialize(cfg Config) (http.Handler, error) {
+	initMu.Lock()
+	defer initMu.Unlock()
+
+	if backend != nil {
+		return backend.HTTPHandler(), nil
+	}
+
+	b, err := observability.New(cfg)
+	if err != nil {
+		return nil, err
+	}
+	backend = b
+
+	if err := createInstruments(); err != nil {
+		backend = nil
+		return nil, err
+	}
+
+	return backend.HTTPHandler(), nil
+}
+
+// Shutdown gracefully shuts down the metrics backend.
+func Shutdown(ctx context.Context) error {
+	initMu.Lock()
+	b := backend
+	backend = nil
+	initMu.Unlock()
+
+	if b != nil {
+		return b.Shutdown(ctx)
+	}
+	return nil
+}
+
+func createInstruments() error {
+	durationBuckets := []float64{0.005, 0.01, 0.025, 0.05, 0.1, 0.25, 0.5, 1, 2.5, 5, 10, 30}
+	sizeBuckets := []float64{512, 1024, 4096, 16384, 65536, 262144, 1048576}
+	sniDurationBuckets := []float64{0.1, 0.5, 1, 2.5, 5, 10, 30, 60, 120}
+
+	b := backend
+
+	newCounter := func(name, desc string, labelNames ...string) (observability.Counter, error) {
+		c, err := b.NewCounter(name, desc, labelNames...)
+		if err != nil {
+			return nil, fmt.Errorf("create counter %q: %w", name, err)
+		}
+		return c, nil
+	}
+
+	newUpDownCounter := func(name, desc string, labelNames ...string) (observability.UpDownCounter, error) {
+		c, err := b.NewUpDownCounter(name, desc, labelNames...)
+		if err != nil {
+			return nil, fmt.Errorf("create updown counter %q: %w", name, err)
+		}
+		return c, nil
+	}
+
+	newInt64Gauge := func(name, desc string, labelNames ...string) (observability.Int64Gauge, error) {
+		g, err := b.NewInt64Gauge(name, desc, labelNames...)
+		if err != nil {
+			return nil, fmt.Errorf("create int64 gauge %q: %w", name, err)
+		}
+		return g, nil
+	}
+
+	newFloat64Gauge := func(name, desc string, labelNames ...string) (observability.Float64Gauge, error) {
+		g, err := b.NewFloat64Gauge(name, desc, labelNames...)
+		if err != nil {
+			return nil, fmt.Errorf("create float64 gauge %q: %w", name, err)
+		}
+		return g, nil
+	}
+
+	newHistogram := func(name, desc string, buckets []float64, labelNames ...string) (observability.Histogram, error) {
+		h, err := b.NewHistogram(name, desc, buckets, labelNames...)
+		if err != nil {
+			return nil, fmt.Errorf("create histogram %q: %w", name, err)
+		}
+		return h, nil
+	}
+
+	var err error
+
+	wgInterfaceUp, err = newInt64Gauge("gerbil_wg_interface_up",
+		"Operational state of a WireGuard interface (1=up, 0=down)", "ifname", "instance")
+	if err != nil {
+		return err
+	}
+	wgPeersTotal, err = newUpDownCounter("gerbil_wg_peers_total",
+		"Total number of configured peers per interface", "ifname")
+	if err != nil {
+		return err
+	}
+	wgPeerConnected, err = newInt64Gauge("gerbil_wg_peer_connected",
+		"Whether a specific peer is connected (1=connected, 0=disconnected)", "ifname", "peer")
+	if err != nil {
+		return err
+	}
+	allowedIPsCount, err = newUpDownCounter("gerbil_allowed_ips_count",
+		"Number of allowed IPs configured per peer", "ifname", "peer")
+	if err != nil {
+		return err
+	}
+	keyRotationTotal, err = newCounter("gerbil_key_rotation_total",
+		"Key rotation events", "ifname", "reason")
+	if err != nil {
+		return err
+	}
+	wgHandshakesTotal, err = newCounter("gerbil_wg_handshakes_total",
+		"Count of handshake attempts with their result status", "ifname", "peer", "result")
+	if err != nil {
+		return err
+	}
+	wgHandshakeLatency, err = newHistogram("gerbil_wg_handshake_latency_seconds",
+		"Distribution of handshake latencies in seconds", durationBuckets, "ifname", "peer")
+	if err != nil {
+		return err
+	}
+	wgPeerRTT, err = newHistogram("gerbil_wg_peer_rtt_seconds",
+		"Observed round-trip time to a peer in seconds", durationBuckets, "ifname", "peer")
+	if err != nil {
+		return err
+	}
+	wgBytesReceived, err = newCounter("gerbil_wg_bytes_received_total",
+		"Number of bytes received from a peer", "ifname", "peer")
+	if err != nil {
+		return err
+	}
+	wgBytesTransmitted, err = newCounter("gerbil_wg_bytes_transmitted_total",
+		"Number of bytes transmitted to a peer", "ifname", "peer")
+	if err != nil {
+		return err
+	}
+	netlinkEventsTotal, err = newCounter("gerbil_netlink_events_total",
+		"Number of netlink events processed", "event_type")
+	if err != nil {
+		return err
+	}
+	netlinkErrorsTotal, err = newCounter("gerbil_netlink_errors_total",
+		"Count of netlink or kernel errors", "component", "error_type")
+	if err != nil {
+		return err
+	}
+	syncDuration, err = newHistogram("gerbil_sync_duration_seconds",
+		"Duration of reconciliation/sync loops in seconds", durationBuckets, "component")
+	if err != nil {
+		return err
+	}
+	workqueueDepth, err = newUpDownCounter("gerbil_workqueue_depth",
+		"Current length of internal work queues", "queue")
+	if err != nil {
+		return err
+	}
+	kernelModuleLoads, err = newCounter("gerbil_kernel_module_loads_total",
+		"Count of kernel module load attempts", "result")
+	if err != nil {
+		return err
+	}
+	firewallRulesApplied, err = newCounter("gerbil_firewall_rules_applied_total",
+		"IPTables/NFT rules applied", "result", "chain")
+	if err != nil {
+		return err
+	}
+	activeSessions, err = newUpDownCounter("gerbil_active_sessions",
+		"Number of active UDP relay sessions", "ifname")
+	if err != nil {
+		return err
+	}
+	activeProxyConnections, err = newUpDownCounter("gerbil_active_proxy_connections",
+		"Active SNI proxy connections")
+	if err != nil {
+		return err
+	}
+	proxyRouteLookups, err = newCounter("gerbil_proxy_route_lookups_total",
+		"Number of route lookups", "result")
+	if err != nil {
+		return err
+	}
+	proxyTLSHandshake, err = newHistogram("gerbil_proxy_tls_handshake_seconds",
+		"TLS handshake duration for SNI proxy in seconds", durationBuckets)
+	if err != nil {
+		return err
+	}
+	proxyBytesTransmitted, err = newCounter("gerbil_proxy_bytes_transmitted_total",
+		"Bytes sent/received by the SNI proxy", "direction")
+	if err != nil {
+		return err
+	}
+	configReloadsTotal, err = newCounter("gerbil_config_reloads_total",
+		"Number of configuration reloads", "result")
+	if err != nil {
+		return err
+	}
+	restartTotal, err = newCounter("gerbil_restart_total",
+		"Process restart count")
+	if err != nil {
+		return err
+	}
+	authFailuresTotal, err = newCounter("gerbil_auth_failures_total",
+		"Count of authentication or peer validation failures", "peer", "reason")
+	if err != nil {
+		return err
+	}
+	aclDeniedTotal, err = newCounter("gerbil_acl_denied_total",
+		"Access control denied events", "ifname", "peer", "policy")
+	if err != nil {
+		return err
+	}
+	certificateExpiryDays, err = newFloat64Gauge("gerbil_certificate_expiry_days",
+		"Days until certificate expiry", "cert_name", "ifname")
+	if err != nil {
+		return err
+	}
+	udpPacketsTotal, err = newCounter("gerbil_udp_packets_total",
+		"Count of UDP packets processed by relay workers", "ifname", "type", "direction")
+	if err != nil {
+		return err
+	}
+	udpPacketSizeBytes, err = newHistogram("gerbil_udp_packet_size_bytes",
+		"Size distribution of packets forwarded through relay", sizeBuckets, "ifname", "type")
+	if err != nil {
+		return err
+	}
+	holePunchEventsTotal, err = newCounter("gerbil_hole_punch_events_total",
+		"Count of hole punch messages processed", "ifname", "result")
+	if err != nil {
+		return err
+	}
+	proxyMappingActive, err = newUpDownCounter("gerbil_proxy_mapping_active",
+		"Number of active proxy mappings", "ifname")
+	if err != nil {
+		return err
+	}
+	relayUDPConnectionsActive, err = newUpDownCounter("gerbil_relay_udp_connections_active",
+		"Number of open per-peer outbound UDP sockets held by the relay connection pool", "ifname")
+	if err != nil {
+		return err
+	}
+	sessionRebuiltTotal, err = newCounter("gerbil_session_rebuilt_total",
+		"Count of sessions rebuilt from communication patterns", "ifname")
+	if err != nil {
+		return err
+	}
+	commPatternActive, err = newUpDownCounter("gerbil_comm_pattern_active",
+		"Number of active communication patterns", "ifname")
+	if err != nil {
+		return err
+	}
+	proxyCleanupRemovedTotal, err = newCounter("gerbil_proxy_cleanup_removed_total",
+		"Count of items removed during cleanup routines", "ifname", "component")
+	if err != nil {
+		return err
+	}
+	proxyConnectionErrorsTotal, err = newCounter("gerbil_proxy_connection_errors_total",
+		"Count of connection errors in proxy operations", "ifname", "error_type")
+	if err != nil {
+		return err
+	}
+	proxyInitialMappingsTotal, err = newInt64Gauge("gerbil_proxy_initial_mappings",
+		"Number of initial proxy mappings loaded", "ifname")
+	if err != nil {
+		return err
+	}
+	proxyMappingUpdatesTotal, err = newCounter("gerbil_proxy_mapping_updates_total",
+		"Count of proxy mapping updates", "ifname")
+	if err != nil {
+		return err
+	}
+	proxyIdleCleanupDuration, err = newHistogram("gerbil_proxy_idle_cleanup_duration_seconds",
+		"Duration of cleanup cycles", durationBuckets, "ifname", "component")
+	if err != nil {
+		return err
+	}
+	sniConnectionsTotal, err = newCounter("gerbil_sni_connections_total",
+		"Count of connections processed by SNI proxy", "result")
+	if err != nil {
+		return err
+	}
+	sniConnectionDuration, err = newHistogram("gerbil_sni_connection_duration_seconds",
+		"Lifetime distribution of proxied TLS connections", sniDurationBuckets)
+	if err != nil {
+		return err
+	}
+	sniActiveConnections, err = newUpDownCounter("gerbil_sni_active_connections",
+		"Number of active SNI tunnels")
+	if err != nil {
+		return err
+	}
+	sniRouteCacheHitsTotal, err = newCounter("gerbil_sni_route_cache_hits_total",
+		"Count of route cache hits and misses", "result")
+	if err != nil {
+		return err
+	}
+	sniRouteAPIRequestsTotal, err = newCounter("gerbil_sni_route_api_requests_total",
+		"Count of route API requests", "result")
+	if err != nil {
+		return err
+	}
+	sniRouteAPILatency, err = newHistogram("gerbil_sni_route_api_latency_seconds",
+		"Distribution of route API call latencies", durationBuckets)
+	if err != nil {
+		return err
+	}
+	sniLocalOverrideTotal, err = newCounter("gerbil_sni_local_override_total",
+		"Count of routes using local overrides", "hit")
+	if err != nil {
+		return err
+	}
+	sniTrustedProxyEventsTotal, err = newCounter("gerbil_sni_trusted_proxy_events_total",
+		"Count of PROXY protocol events", "event")
+	if err != nil {
+		return err
+	}
+	sniProxyProtocolParseErrorsTotal, err = newCounter("gerbil_sni_proxy_protocol_parse_errors_total",
+		"Count of PROXY protocol parse failures")
+	if err != nil {
+		return err
+	}
+	sniDataBytesTotal, err = newCounter("gerbil_sni_data_bytes_total",
+		"Count of bytes proxied through SNI tunnels", "direction")
+	if err != nil {
+		return err
+	}
+	sniTunnelTerminationsTotal, err = newCounter("gerbil_sni_tunnel_terminations_total",
+		"Count of tunnel terminations by reason", "reason")
+	if err != nil {
+		return err
+	}
+	httpRequestsTotal, err = newCounter("gerbil_http_requests_total",
+		"Count of HTTP requests to management API", "endpoint", "method", "status_code")
+	if err != nil {
+		return err
+	}
+	httpRequestDuration, err = newHistogram("gerbil_http_request_duration_seconds",
+		"Distribution of HTTP request handling time", durationBuckets, "endpoint", "method")
+	if err != nil {
+		return err
+	}
+	peerOperationsTotal, err = newCounter("gerbil_peer_operations_total",
+		"Count of peer lifecycle operations", "operation", "result")
+	if err != nil {
+		return err
+	}
+	proxyMappingUpdateRequestsTotal, err = newCounter("gerbil_proxy_mapping_update_requests_total",
+		"Count of proxy mapping update API calls", "result")
+	if err != nil {
+		return err
+	}
+	destinationsUpdateRequestsTotal, err = newCounter("gerbil_destinations_update_requests_total",
+		"Count of destinations update API calls", "result")
+	if err != nil {
+		return err
+	}
+	remoteConfigFetchesTotal, err = newCounter("gerbil_remote_config_fetches_total",
+		"Count of remote configuration fetch attempts", "result")
+	if err != nil {
+		return err
+	}
+	bandwidthReportsTotal, err = newCounter("gerbil_bandwidth_reports_total",
+		"Count of bandwidth report transmissions", "result")
+	if err != nil {
+		return err
+	}
+	peerBandwidthBytesTotal, err = newCounter("gerbil_peer_bandwidth_bytes_total",
+		"Bytes per peer tracked by bandwidth calculation", "peer", "direction")
+	if err != nil {
+		return err
+	}
+	memorySpikeTotal, err = newCounter("gerbil_memory_spike_total",
+		"Count of memory spikes detected", "severity")
+	if err != nil {
+		return err
+	}
+	heapProfilesWrittenTotal, err = newCounter("gerbil_heap_profiles_written_total",
+		"Count of heap profile files generated")
+	if err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func RecordInterfaceUp(ifname, instance string, up bool) {
+	if wgInterfaceUp == nil {
+		return
+	}
+	value := int64(0)
+	if up {
+		value = 1
+	}
+	wgInterfaceUp.Record(context.Background(), value, observability.Labels{"ifname": ifname, "instance": instance})
+}
+
+func RecordPeersTotal(ifname string, delta int64) {
+	if wgPeersTotal == nil {
+		return
+	}
+	wgPeersTotal.Add(context.Background(), delta, observability.Labels{"ifname": ifname})
+}
+
+func RecordPeerConnected(ifname, peer string, connected bool) {
+	if wgPeerConnected == nil {
+		return
+	}
+	value := int64(0)
+	if connected {
+		value = 1
+	}
+	wgPeerConnected.Record(context.Background(), value, observability.Labels{"ifname": ifname, "peer": peer})
+}
+
+func RecordHandshake(ifname, peer, result string) {
+	if wgHandshakesTotal == nil {
+		return
+	}
+	wgHandshakesTotal.Add(context.Background(), 1, observability.Labels{"ifname": ifname, "peer": peer, "result": result})
+}
+
+func RecordHandshakeLatency(ifname, peer string, seconds float64) {
+	if wgHandshakeLatency == nil {
+		return
+	}
+	wgHandshakeLatency.Record(context.Background(), seconds, observability.Labels{"ifname": ifname, "peer": peer})
+}
+
+func RecordPeerRTT(ifname, peer string, seconds float64) {
+	if wgPeerRTT == nil {
+		return
+	}
+	wgPeerRTT.Record(context.Background(), seconds, observability.Labels{"ifname": ifname, "peer": peer})
+}
+
+func RecordBytesReceived(ifname, peer string, bytes int64) {
+	if wgBytesReceived == nil {
+		return
+	}
+	wgBytesReceived.Add(context.Background(), bytes, observability.Labels{"ifname": ifname, "peer": peer})
+}
+
+func RecordBytesTransmitted(ifname, peer string, bytes int64) {
+	if wgBytesTransmitted == nil {
+		return
+	}
+	wgBytesTransmitted.Add(context.Background(), bytes, observability.Labels{"ifname": ifname, "peer": peer})
+}
+
+func RecordAllowedIPsCount(ifname, peer string, delta int64) {
+	if allowedIPsCount == nil {
+		return
+	}
+	allowedIPsCount.Add(context.Background(), delta, observability.Labels{"ifname": ifname, "peer": peer})
+}
+
+func RecordKeyRotation(ifname, reason string) {
+	if keyRotationTotal == nil {
+		return
+	}
+	keyRotationTotal.Add(context.Background(), 1, observability.Labels{"ifname": ifname, "reason": reason})
+}
+
+func RecordNetlinkEvent(eventType string) {
+	if netlinkEventsTotal == nil {
+		return
+	}
+	netlinkEventsTotal.Add(context.Background(), 1, observability.Labels{"event_type": eventType})
+}
+
+func RecordNetlinkError(component, errorType string) {
+	if netlinkErrorsTotal == nil {
+		return
+	}
+	netlinkErrorsTotal.Add(context.Background(), 1, observability.Labels{"component": component, "error_type": errorType})
+}
+
+func RecordSyncDuration(component string, seconds float64) {
+	if syncDuration == nil {
+		return
+	}
+	syncDuration.Record(context.Background(), seconds, observability.Labels{"component": component})
+}
+
+func RecordWorkqueueDepth(queue string, delta int64) {
+	if workqueueDepth == nil {
+		return
+	}
+	workqueueDepth.Add(context.Background(), delta, observability.Labels{"queue": queue})
+}
+
+func RecordKernelModuleLoad(result string) {
+	if kernelModuleLoads == nil {
+		return
+	}
+	kernelModuleLoads.Add(context.Background(), 1, observability.Labels{"result": result})
+}
+
+func RecordFirewallRuleApplied(result, chain string) {
+	if firewallRulesApplied == nil {
+		return
+	}
+	firewallRulesApplied.Add(context.Background(), 1, observability.Labels{"result": result, "chain": chain})
+}
+
+func RecordActiveSession(ifname string, delta int64) {
+	if activeSessions == nil {
+		return
+	}
+	activeSessions.Add(context.Background(), delta, observability.Labels{"ifname": ifname})
+}
+
+func RecordActiveProxyConnection(delta int64) {
+	if activeProxyConnections == nil {
+		return
+	}
+	activeProxyConnections.Add(context.Background(), delta, nil)
+}
+
+func RecordProxyRouteLookup(result string) {
+	if proxyRouteLookups == nil {
+		return
+	}
+	proxyRouteLookups.Add(context.Background(), 1, observability.Labels{"result": result})
+}
+
+func RecordProxyTLSHandshake(seconds float64) {
+	if proxyTLSHandshake == nil {
+		return
+	}
+	proxyTLSHandshake.Record(context.Background(), seconds, nil)
+}
+
+func RecordProxyBytesTransmitted(direction string, bytes int64) {
+	if proxyBytesTransmitted == nil {
+		return
+	}
+	proxyBytesTransmitted.Add(context.Background(), bytes, observability.Labels{"direction": direction})
+}
+
+func RecordConfigReload(result string) {
+	if configReloadsTotal == nil {
+		return
+	}
+	configReloadsTotal.Add(context.Background(), 1, observability.Labels{"result": result})
+}
+
+func RecordRestart() {
+	if restartTotal == nil {
+		return
+	}
+	restartTotal.Add(context.Background(), 1, nil)
+}
+
+func RecordAuthFailure(peer, reason string) {
+	if authFailuresTotal == nil {
+		return
+	}
+	authFailuresTotal.Add(context.Background(), 1, observability.Labels{"peer": peer, "reason": reason})
+}
+
+func RecordACLDenied(ifname, peer, policy string) {
+	if aclDeniedTotal == nil {
+		return
+	}
+	aclDeniedTotal.Add(context.Background(), 1, observability.Labels{"ifname": ifname, "peer": peer, "policy": policy})
+}
+
+func RecordCertificateExpiry(certName, ifname string, days float64) {
+	if certificateExpiryDays == nil {
+		return
+	}
+	certificateExpiryDays.Record(context.Background(), days, observability.Labels{"cert_name": certName, "ifname": ifname})
+}
+
+func RecordUDPPacket(ifname, packetType, direction string) {
+	if udpPacketsTotal == nil {
+		return
+	}
+	udpPacketsTotal.Add(context.Background(), 1, observability.Labels{"ifname": ifname, "type": packetType, "direction": direction})
+}
+
+func RecordUDPPacketSize(ifname, packetType string, bytes float64) {
+	if udpPacketSizeBytes == nil {
+		return
+	}
+	udpPacketSizeBytes.Record(context.Background(), bytes, observability.Labels{"ifname": ifname, "type": packetType})
+}
+
+func RecordHolePunchEvent(ifname, result string) {
+	if holePunchEventsTotal == nil {
+		return
+	}
+	holePunchEventsTotal.Add(context.Background(), 1, observability.Labels{"ifname": ifname, "result": result})
+}
+
+func RecordProxyMapping(ifname string, delta int64) {
+	if proxyMappingActive == nil {
+		return
+	}
+	proxyMappingActive.Add(context.Background(), delta, observability.Labels{"ifname": ifname})
+}
+
+func RecordSession(ifname string, delta int64) {
+	if activeSessions == nil {
+		return
+	}
+	activeSessions.Add(context.Background(), delta, observability.Labels{"ifname": ifname})
+}
+
+func RecordUDPConnection(ifname string, delta int64) {
+	if relayUDPConnectionsActive == nil {
+		return
+	}
+	relayUDPConnectionsActive.Add(context.Background(), delta, observability.Labels{"ifname": ifname})
+}
+
+func RecordSessionRebuilt(ifname string) {
+	if sessionRebuiltTotal == nil {
+		return
+	}
+	sessionRebuiltTotal.Add(context.Background(), 1, observability.Labels{"ifname": ifname})
+}
+
+func RecordCommPattern(ifname string, delta int64) {
+	if commPatternActive == nil {
+		return
+	}
+	commPatternActive.Add(context.Background(), delta, observability.Labels{"ifname": ifname})
+}
+
+func RecordProxyCleanupRemoved(ifname, component string, count int64) {
+	if proxyCleanupRemovedTotal == nil {
+		return
+	}
+	proxyCleanupRemovedTotal.Add(context.Background(), count, observability.Labels{"ifname": ifname, "component": component})
+}
+
+func RecordProxyConnectionError(ifname, errorType string) {
+	if proxyConnectionErrorsTotal == nil {
+		return
+	}
+	proxyConnectionErrorsTotal.Add(context.Background(), 1, observability.Labels{"ifname": ifname, "error_type": errorType})
+}
+
+func RecordProxyInitialMappings(ifname string, count int64) {
+	if proxyInitialMappingsTotal == nil {
+		return
+	}
+	proxyInitialMappingsTotal.Record(context.Background(), count, observability.Labels{"ifname": ifname})
+}
+
+func RecordProxyMappingUpdate(ifname string) {
+	if proxyMappingUpdatesTotal == nil {
+		return
+	}
+	proxyMappingUpdatesTotal.Add(context.Background(), 1, observability.Labels{"ifname": ifname})
+}
+
+func RecordProxyIdleCleanupDuration(ifname, component string, seconds float64) {
+	if proxyIdleCleanupDuration == nil {
+		return
+	}
+	proxyIdleCleanupDuration.Record(context.Background(), seconds, observability.Labels{"ifname": ifname, "component": component})
+}
+
+func RecordSNIConnection(result string) {
+	if sniConnectionsTotal == nil {
+		return
+	}
+	sniConnectionsTotal.Add(context.Background(), 1, observability.Labels{"result": result})
+}
+
+func RecordSNIConnectionDuration(seconds float64) {
+	if sniConnectionDuration == nil {
+		return
+	}
+	sniConnectionDuration.Record(context.Background(), seconds, nil)
+}
+
+func RecordSNIActiveConnection(delta int64) {
+	if sniActiveConnections == nil {
+		return
+	}
+	sniActiveConnections.Add(context.Background(), delta, nil)
+}
+
+func RecordSNIRouteCacheHit(result string) {
+	if sniRouteCacheHitsTotal == nil {
+		return
+	}
+	sniRouteCacheHitsTotal.Add(context.Background(), 1, observability.Labels{"result": result})
+}
+
+func RecordSNIRouteAPIRequest(result string) {
+	if sniRouteAPIRequestsTotal == nil {
+		return
+	}
+	sniRouteAPIRequestsTotal.Add(context.Background(), 1, observability.Labels{"result": result})
+}
+
+func RecordSNIRouteAPILatency(seconds float64) {
+	if sniRouteAPILatency == nil {
+		return
+	}
+	sniRouteAPILatency.Record(context.Background(), seconds, nil)
+}
+
+func RecordSNILocalOverride(hit string) {
+	if sniLocalOverrideTotal == nil {
+		return
+	}
+	sniLocalOverrideTotal.Add(context.Background(), 1, observability.Labels{"hit": hit})
+}
+
+func RecordSNITrustedProxyEvent(event string) {
+	if sniTrustedProxyEventsTotal == nil {
+		return
+	}
+	sniTrustedProxyEventsTotal.Add(context.Background(), 1, observability.Labels{"event": event})
+}
+
+func RecordSNIProxyProtocolParseError() {
+	if sniProxyProtocolParseErrorsTotal == nil {
+		return
+	}
+	sniProxyProtocolParseErrorsTotal.Add(context.Background(), 1, nil)
+}
+
+func RecordSNIDataBytes(direction string, bytes int64) {
+	if sniDataBytesTotal == nil {
+		return
+	}
+	sniDataBytesTotal.Add(context.Background(), bytes, observability.Labels{"direction": direction})
+}
+
+func RecordSNITunnelTermination(reason string) {
+	if sniTunnelTerminationsTotal == nil {
+		return
+	}
+	sniTunnelTerminationsTotal.Add(context.Background(), 1, observability.Labels{"reason": reason})
+}
+
+func RecordHTTPRequest(endpoint, method, statusCode string) {
+	if httpRequestsTotal == nil {
+		return
+	}
+	httpRequestsTotal.Add(context.Background(), 1, observability.Labels{"endpoint": endpoint, "method": method, "status_code": statusCode})
+}
+
+func RecordHTTPRequestDuration(endpoint, method string, seconds float64) {
+	if httpRequestDuration == nil {
+		return
+	}
+	httpRequestDuration.Record(context.Background(), seconds, observability.Labels{"endpoint": endpoint, "method": method})
+}
+
+func RecordPeerOperation(operation, result string) {
+	if peerOperationsTotal == nil {
+		return
+	}
+	peerOperationsTotal.Add(context.Background(), 1, observability.Labels{"operation": operation, "result": result})
+}
+
+func RecordProxyMappingUpdateRequest(result string) {
+	if proxyMappingUpdateRequestsTotal == nil {
+		return
+	}
+	proxyMappingUpdateRequestsTotal.Add(context.Background(), 1, observability.Labels{"result": result})
+}
+
+func RecordDestinationsUpdateRequest(result string) {
+	if destinationsUpdateRequestsTotal == nil {
+		return
+	}
+	destinationsUpdateRequestsTotal.Add(context.Background(), 1, observability.Labels{"result": result})
+}
+
+func RecordRemoteConfigFetch(result string) {
+	if remoteConfigFetchesTotal == nil {
+		return
+	}
+	remoteConfigFetchesTotal.Add(context.Background(), 1, observability.Labels{"result": result})
+}
+
+func RecordBandwidthReport(result string) {
+	if bandwidthReportsTotal == nil {
+		return
+	}
+	bandwidthReportsTotal.Add(context.Background(), 1, observability.Labels{"result": result})
+}
+
+func RecordPeerBandwidthBytes(peer, direction string, bytes int64) {
+	if peerBandwidthBytesTotal == nil {
+		return
+	}
+	peerBandwidthBytesTotal.Add(context.Background(), bytes, observability.Labels{"peer": peer, "direction": direction})
+}
+
+func RecordMemorySpike(severity string) {
+	if memorySpikeTotal == nil {
+		return
+	}
+	memorySpikeTotal.Add(context.Background(), 1, observability.Labels{"severity": severity})
+}
+
+func RecordHeapProfileWritten() {
+	if heapProfilesWrittenTotal == nil {
+		return
+	}
+	heapProfilesWrittenTotal.Add(context.Background(), 1, nil)
+}

internal/metrics/metrics_test.go

@@ -0,0 +1,262 @@
+package metrics_test
+
+import (
+	"context"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/fosrl/gerbil/internal/metrics"
+	"github.com/fosrl/gerbil/internal/observability"
+)
+
+const exampleHostname = "example.com"
+
+func initPrometheus(t *testing.T) http.Handler {
+	t.Helper()
+	cfg := metrics.DefaultConfig()
+	cfg.Enabled = true
+	cfg.Backend = "prometheus"
+	cfg.Prometheus.Path = "/metrics"
+
+	h, err := metrics.Initialize(cfg)
+	if err != nil {
+		t.Fatalf("Initialize failed: %v", err)
+	}
+	t.Cleanup(func() {
+		metrics.Shutdown(context.Background()) //nolint:errcheck
+	})
+	return h
+}
+
+func initNoop(t *testing.T) {
+	t.Helper()
+	cfg := metrics.DefaultConfig()
+	cfg.Enabled = false
+	_, err := metrics.Initialize(cfg)
+	if err != nil {
+		t.Fatalf("Initialize noop failed: %v", err)
+	}
+	t.Cleanup(func() {
+		metrics.Shutdown(context.Background()) //nolint:errcheck
+	})
+}
+
+func scrape(t *testing.T, h http.Handler) string {
+	t.Helper()
+	req := httptest.NewRequest(http.MethodGet, "/metrics", http.NoBody)
+	rr := httptest.NewRecorder()
+	h.ServeHTTP(rr, req)
+	if rr.Code != http.StatusOK {
+		t.Fatalf("scrape returned %d", rr.Code)
+	}
+	b, _ := io.ReadAll(rr.Body)
+	return string(b)
+}
+
+func assertContains(t *testing.T, body, substr string) {
+	t.Helper()
+	if !strings.Contains(body, substr) {
+		t.Errorf("expected %q in output\nbody:\n%s", substr, body)
+	}
+}
+
+// --- Tests ---
+
+func TestInitializePrometheus(t *testing.T) {
+	h := initPrometheus(t)
+	if h == nil {
+		t.Error("expected non-nil HTTP handler for prometheus backend")
+	}
+}
+
+func TestInitializeNoop(t *testing.T) {
+	initNoop(t)
+	// All Record* functions must not panic when noop backend is active.
+	metrics.RecordRestart()
+	metrics.RecordHTTPRequest("/test", "GET", "200")
+	metrics.RecordSNIConnection("accepted")
+	metrics.RecordPeersTotal("wg0", 1)
+}
+
+func TestDefaultConfig(t *testing.T) {
+	cfg := metrics.DefaultConfig()
+	if cfg.Backend != "prometheus" {
+		t.Errorf("expected prometheus default backend, got %q", cfg.Backend)
+	}
+}
+
+func TestShutdownNoInit(t *testing.T) {
+	// Ensure a known clean global state before testing no-init shutdown behavior.
+	_ = metrics.Shutdown(context.Background())
+
+	// Shutdown without Initialize should not panic or error.
+	if err := metrics.Shutdown(context.Background()); err != nil {
+		t.Errorf("unexpected error: %v", err)
+	}
+}
+
+func TestRecordHTTPRequest(t *testing.T) {
+	h := initPrometheus(t)
+	metrics.RecordHTTPRequest("/peers", "POST", "201")
+	body := scrape(t, h)
+	assertContains(t, body, "gerbil_http_requests_total")
+}
+
+func TestRecordHTTPRequestDuration(t *testing.T) {
+	h := initPrometheus(t)
+	metrics.RecordHTTPRequestDuration("/peers", "POST", 0.05)
+	body := scrape(t, h)
+	assertContains(t, body, "gerbil_http_request_duration_seconds")
+}
+
+func TestRecordInterfaceUp(t *testing.T) {
+	h := initPrometheus(t)
+	metrics.RecordInterfaceUp("wg0", "host1", true)
+	metrics.RecordInterfaceUp("wg0", "host1", false)
+	body := scrape(t, h)
+	assertContains(t, body, "gerbil_wg_interface_up")
+}
+
+func TestRecordPeersTotal(t *testing.T) {
+	h := initPrometheus(t)
+	metrics.RecordPeersTotal("wg0", 3)
+	body := scrape(t, h)
+	assertContains(t, body, "gerbil_wg_peers_total")
+}
+
+func TestRecordBytesReceivedTransmitted(t *testing.T) {
+	h := initPrometheus(t)
+	metrics.RecordBytesReceived("wg0", "peer1", 1024)
+	metrics.RecordBytesTransmitted("wg0", "peer1", 512)
+	body := scrape(t, h)
+	assertContains(t, body, "gerbil_wg_bytes_received_total")
+	assertContains(t, body, "gerbil_wg_bytes_transmitted_total")
+}
+
+func TestRecordSNI(t *testing.T) {
+	h := initPrometheus(t)
+	metrics.RecordSNIConnection("accepted")
+	metrics.RecordSNIActiveConnection(1)
+	metrics.RecordSNIConnectionDuration(1.5)
+	metrics.RecordSNIRouteCacheHit("hit")
+	metrics.RecordSNIRouteAPIRequest("success")
+	metrics.RecordSNIRouteAPILatency(0.01)
+	metrics.RecordSNILocalOverride("yes")
+	metrics.RecordSNITrustedProxyEvent("proxy_protocol_parsed")
+	metrics.RecordSNIProxyProtocolParseError()
+	metrics.RecordSNIDataBytes("client_to_target", 2048)
+	metrics.RecordSNITunnelTermination("eof")
+	body := scrape(t, h)
+	assertContains(t, body, "gerbil_sni_connections_total")
+	assertContains(t, body, "gerbil_sni_active_connections")
+}
+
+func TestRecordRelay(t *testing.T) {
+	h := initPrometheus(t)
+	metrics.RecordUDPPacket("relay", "data", "in")
+	metrics.RecordUDPPacketSize("relay", "data", 256)
+	metrics.RecordHolePunchEvent("relay", "success")
+	metrics.RecordProxyMapping("relay", 1)
+	metrics.RecordSession("relay", 1)
+	metrics.RecordSessionRebuilt("relay")
+	metrics.RecordCommPattern("relay", 1)
+	metrics.RecordProxyCleanupRemoved("relay", "session", 2)
+	metrics.RecordProxyConnectionError("relay", "dial_udp")
+	metrics.RecordProxyInitialMappings("relay", 5)
+	metrics.RecordProxyMappingUpdate("relay")
+	metrics.RecordProxyIdleCleanupDuration("relay", "conn", 0.1)
+	body := scrape(t, h)
+	assertContains(t, body, "gerbil_udp_packets_total")
+	assertContains(t, body, "gerbil_proxy_mapping_active")
+	assertContains(t, body, "gerbil_active_sessions")
+}
+
+func TestRecordWireGuard(t *testing.T) {
+	h := initPrometheus(t)
+	metrics.RecordHandshake("wg0", "peer1", "success")
+	metrics.RecordHandshakeLatency("wg0", "peer1", 0.02)
+	metrics.RecordPeerRTT("wg0", "peer1", 0.005)
+	metrics.RecordPeerConnected("wg0", "peer1", true)
+	metrics.RecordAllowedIPsCount("wg0", "peer1", 2)
+	metrics.RecordKeyRotation("wg0", "scheduled")
+	body := scrape(t, h)
+	assertContains(t, body, "gerbil_wg_handshakes_total")
+	assertContains(t, body, "gerbil_wg_peer_connected")
+}
+
+func TestRecordHousekeeping(t *testing.T) {
+	h := initPrometheus(t)
+	metrics.RecordRemoteConfigFetch("success")
+	metrics.RecordBandwidthReport("success")
+	metrics.RecordPeerBandwidthBytes("peer1", "rx", 512)
+	metrics.RecordMemorySpike("warning")
+	metrics.RecordHeapProfileWritten()
+	body := scrape(t, h)
+	assertContains(t, body, "gerbil_remote_config_fetches_total")
+	assertContains(t, body, "gerbil_memory_spike_total")
+}
+
+func TestRecordOperational(t *testing.T) {
+	h := initPrometheus(t)
+	metrics.RecordConfigReload("success")
+	metrics.RecordRestart()
+	metrics.RecordAuthFailure("peer1", "bad_key")
+	metrics.RecordACLDenied("wg0", "peer1", "default-deny")
+	metrics.RecordCertificateExpiry(exampleHostname, "wg0", 90.0)
+	body := scrape(t, h)
+	assertContains(t, body, "gerbil_config_reloads_total")
+	assertContains(t, body, "gerbil_restart_total")
+}
+
+func TestRecordNetlink(t *testing.T) {
+	h := initPrometheus(t)
+	metrics.RecordNetlinkEvent("link_up")
+	metrics.RecordNetlinkError("wg", "timeout")
+	metrics.RecordSyncDuration("config", 0.1)
+	metrics.RecordWorkqueueDepth("main", 3)
+	metrics.RecordKernelModuleLoad("success")
+	metrics.RecordFirewallRuleApplied("success", "INPUT")
+	metrics.RecordActiveSession("wg0", 1)
+	metrics.RecordActiveProxyConnection(1)
+	metrics.RecordProxyRouteLookup("hit")
+	metrics.RecordProxyTLSHandshake(0.05)
+	metrics.RecordProxyBytesTransmitted("tx", 1024)
+	body := scrape(t, h)
+	assertContains(t, body, "gerbil_netlink_events_total")
+	assertContains(t, body, "gerbil_active_sessions")
+}
+
+func TestRecordPeerOperation(t *testing.T) {
+	h := initPrometheus(t)
+	metrics.RecordPeerOperation("add", "success")
+	metrics.RecordProxyMappingUpdateRequest("success")
+	metrics.RecordDestinationsUpdateRequest("success")
+	body := scrape(t, h)
+	assertContains(t, body, "gerbil_peer_operations_total")
+}
+
+func TestInitializeInvalidBackend(t *testing.T) {
+	cfg := observability.MetricsConfig{Enabled: true, Backend: "invalid"}
+	_, err := metrics.Initialize(cfg)
+	if err == nil {
+		t.Error("expected error for invalid backend")
+	}
+}
+
+func TestInitializeBackendNone(t *testing.T) {
+	cfg := metrics.DefaultConfig()
+	cfg.Backend = "none"
+	h, err := metrics.Initialize(cfg)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if h != nil {
+		t.Error("none backend should return nil handler")
+	}
+	// All Record* calls should be noop
+	metrics.RecordRestart()
+	metrics.Shutdown(context.Background()) //nolint:errcheck
+}

internal/observability/config.go

@@ -0,0 +1,129 @@
+// Package observability provides a backend-neutral metrics abstraction for Gerbil.
+//
+// Exactly one metrics backend may be enabled at runtime:
+//   - "prometheus" – native Prometheus client; exposes /metrics (no OTel SDK required)
+//   - "otel"       – OpenTelemetry metrics pushed via OTLP (gRPC or HTTP)
+//   - "none"       – metrics disabled; a safe noop implementation is used
+//
+// Future OTel tracing and logging can be added to this package alongside the
+// existing otel sub-package without touching the Prometheus-native path.
+package observability
+
+import (
+	"fmt"
+	"time"
+)
+
+// MetricsConfig is the top-level metrics configuration.
+type MetricsConfig struct {
+	// Enabled controls whether any metrics backend is started.
+	// When false the noop backend is used regardless of Backend.
+	Enabled bool
+
+	// Backend selects the active backend: "prometheus", "otel", or "none".
+	Backend string
+
+	// Prometheus holds settings used only by the Prometheus-native backend.
+	Prometheus PrometheusConfig
+
+	// OTel holds settings used only by the OTel backend.
+	OTel OTelConfig
+
+	// ServiceName is propagated to OTel resource attributes.
+	ServiceName string
+
+	// ServiceVersion is propagated to OTel resource attributes.
+	ServiceVersion string
+
+	// DeploymentEnvironment is an optional OTel resource attribute.
+	DeploymentEnvironment string
+}
+
+// PrometheusConfig holds Prometheus-native backend settings.
+type PrometheusConfig struct {
+	// Path is the HTTP path to expose the /metrics endpoint.
+	// Defaults to "/metrics".
+	Path string
+}
+
+// OTelConfig holds OpenTelemetry backend settings.
+type OTelConfig struct {
+	// Protocol is the OTLP transport: "grpc" (default) or "http".
+	Protocol string
+
+	// Endpoint is the OTLP collector address (e.g. "localhost:4317").
+	Endpoint string
+
+	// Insecure disables TLS for the OTLP connection.
+	Insecure bool
+
+	// ExportInterval is how often metrics are pushed to the collector.
+	// Defaults to 60 s.
+	ExportInterval time.Duration
+
+	// Timeout bounds OTLP exporter construction calls.
+	// Defaults to 10 s.
+	Timeout time.Duration
+}
+
+// DefaultMetricsConfig returns a MetricsConfig with sensible defaults.
+func DefaultMetricsConfig() MetricsConfig {
+	return MetricsConfig{
+		Enabled: true,
+		Backend: "prometheus",
+		Prometheus: PrometheusConfig{
+			Path: "/metrics",
+		},
+		OTel: OTelConfig{
+			Protocol:       "grpc",
+			Endpoint:       "localhost:4317",
+			Insecure:       true,
+			ExportInterval: 60 * time.Second,
+			Timeout:        10 * time.Second,
+		},
+		ServiceName:    "gerbil",
+		ServiceVersion: "1.0.0",
+	}
+}
+
+// Validate checks the configuration for logical errors.
+func (c *MetricsConfig) Validate() error {
+	if !c.Enabled {
+		return nil
+	}
+
+	switch c.Backend {
+	case "prometheus", "none":
+		// valid
+	case "":
+		return fmt.Errorf("metrics: enabled requires a non-empty backend")
+	case "otel":
+		if c.OTel.Endpoint == "" {
+			return fmt.Errorf("metrics: backend=otel requires a non-empty OTel endpoint")
+		}
+		if c.OTel.Protocol != "grpc" && c.OTel.Protocol != "http" {
+			return fmt.Errorf("metrics: otel protocol must be \"grpc\" or \"http\", got %q", c.OTel.Protocol)
+		}
+		if c.OTel.ExportInterval <= 0 {
+			return fmt.Errorf("metrics: otel export interval must be positive")
+		}
+		if c.OTel.Timeout <= 0 {
+			return fmt.Errorf("metrics: otel timeout must be positive")
+		}
+	default:
+		return fmt.Errorf("metrics: unknown backend %q (must be \"prometheus\", \"otel\", or \"none\")", c.Backend)
+	}
+
+	return nil
+}
+
+// effectiveBackend resolves the backend string, treating "" and "none" as noop.
+func (c *MetricsConfig) effectiveBackend() string {
+	if !c.Enabled {
+		return "none"
+	}
+	if c.Backend == "" {
+		return "none"
+	}
+	return c.Backend
+}

internal/observability/metrics.go

@@ -0,0 +1,153 @@
+package observability
+
+import (
+	"context"
+	"fmt"
+	"net/http"
+
+	obsotel "github.com/fosrl/gerbil/internal/observability/otel"
+	obsprom "github.com/fosrl/gerbil/internal/observability/prometheus"
+)
+
+// Labels is a set of key-value pairs attached to a metric observation.
+// Use only stable, bounded-cardinality label values.
+type Labels = map[string]string
+
+// Counter is a monotonically increasing instrument.
+type Counter interface {
+	Add(ctx context.Context, value int64, labels Labels)
+}
+
+// UpDownCounter is a bidirectional integer instrument (can go up or down).
+type UpDownCounter interface {
+	Add(ctx context.Context, value int64, labels Labels)
+}
+
+// Int64Gauge records a snapshot integer value.
+type Int64Gauge interface {
+	Record(ctx context.Context, value int64, labels Labels)
+}
+
+// Float64Gauge records a snapshot float value.
+type Float64Gauge interface {
+	Record(ctx context.Context, value float64, labels Labels)
+}
+
+// Histogram records a distribution of values.
+type Histogram interface {
+	Record(ctx context.Context, value float64, labels Labels)
+}
+
+// Backend is the single interface that each metrics implementation must satisfy.
+// Application code must not import backend-specific packages (prometheus, otel).
+type Backend interface {
+	// NewCounter creates a counter metric.
+	// labelNames declares the set of label keys that will be passed at observation time.
+	NewCounter(name, desc string, labelNames ...string) (Counter, error)
+
+	// NewUpDownCounter creates an up-down counter metric.
+	NewUpDownCounter(name, desc string, labelNames ...string) (UpDownCounter, error)
+
+	// NewInt64Gauge creates an integer gauge metric.
+	NewInt64Gauge(name, desc string, labelNames ...string) (Int64Gauge, error)
+
+	// NewFloat64Gauge creates a float gauge metric.
+	NewFloat64Gauge(name, desc string, labelNames ...string) (Float64Gauge, error)
+
+	// NewHistogram creates a histogram metric.
+	// buckets are the explicit upper-bound bucket boundaries.
+	NewHistogram(name, desc string, buckets []float64, labelNames ...string) (Histogram, error)
+
+	// HTTPHandler returns the /metrics HTTP handler.
+	// Implementations that do not expose an HTTP endpoint return nil.
+	HTTPHandler() http.Handler
+
+	// Shutdown performs a graceful flush / shutdown of the backend.
+	Shutdown(ctx context.Context) error
+}
+
+// New creates the backend selected by cfg and returns it.
+// Exactly one backend is created; the selection is mutually exclusive.
+func New(cfg MetricsConfig) (Backend, error) {
+	if err := cfg.Validate(); err != nil {
+		return nil, err
+	}
+
+	switch cfg.effectiveBackend() {
+	case "prometheus":
+		b, err := obsprom.New(obsprom.Config{
+			Path: cfg.Prometheus.Path,
+		})
+		if err != nil {
+			return nil, err
+		}
+		return &promAdapter{b: b}, nil
+	case "otel":
+		b, err := obsotel.New(obsotel.Config{
+			Protocol:              cfg.OTel.Protocol,
+			Endpoint:              cfg.OTel.Endpoint,
+			Insecure:              cfg.OTel.Insecure,
+			ExportInterval:        cfg.OTel.ExportInterval,
+			Timeout:               cfg.OTel.Timeout,
+			ServiceName:           cfg.ServiceName,
+			ServiceVersion:        cfg.ServiceVersion,
+			DeploymentEnvironment: cfg.DeploymentEnvironment,
+		})
+		if err != nil {
+			return nil, err
+		}
+		return &otelAdapter{b: b}, nil
+	case "none":
+		return &NoopBackend{}, nil
+	default:
+		return nil, fmt.Errorf("observability: unknown backend %q", cfg.effectiveBackend())
+	}
+}
+
+// promAdapter wraps obsprom.Backend to implement the observability.Backend interface.
+// The concrete instrument types from the prometheus sub-package satisfy the instrument
+// interfaces via Go's structural (duck) typing without importing this package.
+type promAdapter struct {
+	b *obsprom.Backend
+}
+
+func (a *promAdapter) NewCounter(name, desc string, labelNames ...string) (Counter, error) {
+	return a.b.NewCounter(name, desc, labelNames...)
+}
+func (a *promAdapter) NewUpDownCounter(name, desc string, labelNames ...string) (UpDownCounter, error) {
+	return a.b.NewUpDownCounter(name, desc, labelNames...)
+}
+func (a *promAdapter) NewInt64Gauge(name, desc string, labelNames ...string) (Int64Gauge, error) {
+	return a.b.NewInt64Gauge(name, desc, labelNames...)
+}
+func (a *promAdapter) NewFloat64Gauge(name, desc string, labelNames ...string) (Float64Gauge, error) {
+	return a.b.NewFloat64Gauge(name, desc, labelNames...)
+}
+func (a *promAdapter) NewHistogram(name, desc string, buckets []float64, labelNames ...string) (Histogram, error) {
+	return a.b.NewHistogram(name, desc, buckets, labelNames...)
+}
+func (a *promAdapter) HTTPHandler() http.Handler          { return a.b.HTTPHandler() }
+func (a *promAdapter) Shutdown(ctx context.Context) error { return a.b.Shutdown(ctx) }
+
+// otelAdapter wraps obsotel.Backend to implement the observability.Backend interface.
+type otelAdapter struct {
+	b *obsotel.Backend
+}
+
+func (a *otelAdapter) NewCounter(name, desc string, labelNames ...string) (Counter, error) {
+	return a.b.NewCounter(name, desc, labelNames...)
+}
+func (a *otelAdapter) NewUpDownCounter(name, desc string, labelNames ...string) (UpDownCounter, error) {
+	return a.b.NewUpDownCounter(name, desc, labelNames...)
+}
+func (a *otelAdapter) NewInt64Gauge(name, desc string, labelNames ...string) (Int64Gauge, error) {
+	return a.b.NewInt64Gauge(name, desc, labelNames...)
+}
+func (a *otelAdapter) NewFloat64Gauge(name, desc string, labelNames ...string) (Float64Gauge, error) {
+	return a.b.NewFloat64Gauge(name, desc, labelNames...)
+}
+func (a *otelAdapter) NewHistogram(name, desc string, buckets []float64, labelNames ...string) (Histogram, error) {
+	return a.b.NewHistogram(name, desc, buckets, labelNames...)
+}
+func (a *otelAdapter) HTTPHandler() http.Handler          { return a.b.HTTPHandler() }
+func (a *otelAdapter) Shutdown(ctx context.Context) error { return a.b.Shutdown(ctx) }

internal/observability/metrics_test.go

@@ -0,0 +1,263 @@
+package observability_test
+
+import (
+	"context"
+	"net"
+	"os"
+	"testing"
+	"time"
+
+	"github.com/fosrl/gerbil/internal/observability"
+)
+
+const (
+	defaultMetricsPath = "/metrics"
+	otelGRPCEndpoint   = "localhost:4317"
+	errUnexpectedFmt   = "unexpected error: %v"
+)
+
+func TestDefaultMetricsConfig(t *testing.T) {
+	cfg := observability.DefaultMetricsConfig()
+	if !cfg.Enabled {
+		t.Error("default config should have Enabled=true")
+	}
+	if cfg.Backend != "prometheus" {
+		t.Errorf("default backend should be prometheus, got %q", cfg.Backend)
+	}
+	if cfg.Prometheus.Path != defaultMetricsPath {
+		t.Errorf("default prometheus path should be %s, got %q", defaultMetricsPath, cfg.Prometheus.Path)
+	}
+	if cfg.OTel.Protocol != "grpc" {
+		t.Errorf("default otel protocol should be grpc, got %q", cfg.OTel.Protocol)
+	}
+	if cfg.OTel.ExportInterval != 60*time.Second {
+		t.Errorf("default otel export interval should be 60s, got %v", cfg.OTel.ExportInterval)
+	}
+}
+func TestValidateValidConfigs(t *testing.T) {
+	tests := []struct {
+		name string
+		cfg  observability.MetricsConfig
+	}{
+		{name: "disabled", cfg: observability.MetricsConfig{Enabled: false}},
+		{name: "backend none", cfg: observability.MetricsConfig{Enabled: true, Backend: "none"}},
+		{name: "prometheus", cfg: observability.MetricsConfig{Enabled: true, Backend: "prometheus"}},
+		{
+			name: "otel grpc",
+			cfg: observability.MetricsConfig{
+				Enabled: true, Backend: "otel",
+				OTel: observability.OTelConfig{Protocol: "grpc", Endpoint: otelGRPCEndpoint, ExportInterval: 10 * time.Second, Timeout: 2 * time.Second},
+			},
+		},
+		{
+			name: "otel http",
+			cfg: observability.MetricsConfig{
+				Enabled: true, Backend: "otel",
+				OTel: observability.OTelConfig{Protocol: "http", Endpoint: "localhost:4318", ExportInterval: 30 * time.Second, Timeout: 2 * time.Second},
+			},
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if err := tt.cfg.Validate(); err != nil {
+				t.Errorf("unexpected validation error: %v", err)
+			}
+		})
+	}
+}
+
+func TestValidateInvalidConfigs(t *testing.T) {
+	tests := []struct {
+		name string
+		cfg  observability.MetricsConfig
+	}{
+		{name: "unknown backend", cfg: observability.MetricsConfig{Enabled: true, Backend: "datadog"}},
+		{
+			name: "backend empty while enabled",
+			cfg:  observability.MetricsConfig{Enabled: true, Backend: ""},
+		},
+		{
+			name: "otel missing endpoint",
+			cfg: observability.MetricsConfig{
+				Enabled: true, Backend: "otel",
+				OTel: observability.OTelConfig{Protocol: "grpc", Endpoint: "", ExportInterval: 10 * time.Second, Timeout: 2 * time.Second},
+			},
+		},
+		{
+			name: "otel invalid protocol",
+			cfg: observability.MetricsConfig{
+				Enabled: true, Backend: "otel",
+				OTel: observability.OTelConfig{Protocol: "tcp", Endpoint: otelGRPCEndpoint, ExportInterval: 10 * time.Second, Timeout: 2 * time.Second},
+			},
+		},
+		{
+			name: "otel zero interval",
+			cfg: observability.MetricsConfig{
+				Enabled: true, Backend: "otel",
+				OTel: observability.OTelConfig{Protocol: "grpc", Endpoint: otelGRPCEndpoint, ExportInterval: 0, Timeout: 2 * time.Second},
+			},
+		},
+		{
+			name: "otel zero timeout",
+			cfg: observability.MetricsConfig{
+				Enabled: true, Backend: "otel",
+				OTel: observability.OTelConfig{Protocol: "grpc", Endpoint: otelGRPCEndpoint, ExportInterval: 10 * time.Second, Timeout: 0},
+			},
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if err := tt.cfg.Validate(); err == nil {
+				t.Error("expected validation error but got nil")
+			}
+		})
+	}
+}
+
+func TestNewNoopBackend(t *testing.T) {
+	b, err := observability.New(observability.MetricsConfig{Enabled: false})
+	if err != nil {
+		t.Fatalf(errUnexpectedFmt, err)
+	}
+	if b.HTTPHandler() != nil {
+		t.Error("noop backend HTTPHandler should return nil")
+	}
+}
+
+func TestNewNoneBackend(t *testing.T) {
+	b, err := observability.New(observability.MetricsConfig{Enabled: true, Backend: "none"})
+	if err != nil {
+		t.Fatalf(errUnexpectedFmt, err)
+	}
+	if b.HTTPHandler() != nil {
+		t.Error("none backend HTTPHandler should return nil")
+	}
+}
+
+func TestNewPrometheusBackend(t *testing.T) {
+	cfg := observability.MetricsConfig{
+		Enabled: true, Backend: "prometheus",
+		Prometheus: observability.PrometheusConfig{Path: defaultMetricsPath},
+	}
+	b, err := observability.New(cfg)
+	if err != nil {
+		t.Fatalf(errUnexpectedFmt, err)
+	}
+	if b.HTTPHandler() == nil {
+		t.Error("prometheus backend HTTPHandler should not be nil")
+	}
+	if err := b.Shutdown(context.Background()); err != nil {
+		t.Errorf("prometheus shutdown error: %v", err)
+	}
+}
+
+func TestNewInvalidBackend(t *testing.T) {
+	_, err := observability.New(observability.MetricsConfig{Enabled: true, Backend: "invalid"})
+	if err == nil {
+		t.Error("expected error for invalid backend")
+	}
+}
+
+func TestPrometheusAdapterAllInstruments(t *testing.T) {
+	b, err := observability.New(observability.MetricsConfig{
+		Enabled: true, Backend: "prometheus",
+		Prometheus: observability.PrometheusConfig{Path: defaultMetricsPath},
+	})
+	if err != nil {
+		t.Fatalf("failed to create backend: %v", err)
+	}
+	ctx := context.Background()
+	labels := observability.Labels{"k": "v"}
+
+	c, err := b.NewCounter("prom_adapter_counter_total", "desc", "k")
+	if err != nil {
+		t.Fatalf("NewCounter error: %v", err)
+	}
+	u, err := b.NewUpDownCounter("prom_adapter_updown", "desc", "k")
+	if err != nil {
+		t.Fatalf("NewUpDownCounter error: %v", err)
+	}
+	ig, err := b.NewInt64Gauge("prom_adapter_int_gauge", "desc", "k")
+	if err != nil {
+		t.Fatalf("NewInt64Gauge error: %v", err)
+	}
+	fg, err := b.NewFloat64Gauge("prom_adapter_float_gauge", "desc", "k")
+	if err != nil {
+		t.Fatalf("NewFloat64Gauge error: %v", err)
+	}
+	h, err := b.NewHistogram("prom_adapter_histogram", "desc", []float64{0.1, 1.0}, "k")
+	if err != nil {
+		t.Fatalf("NewHistogram error: %v", err)
+	}
+
+	c.Add(ctx, 1, labels)
+	u.Add(ctx, 2, labels)
+	ig.Record(ctx, 99, labels)
+	fg.Record(ctx, 1.23, labels)
+	h.Record(ctx, 0.5, labels)
+
+	if b.HTTPHandler() == nil {
+		t.Error("prometheus adapter HTTPHandler should not be nil")
+	}
+	if err := b.Shutdown(ctx); err != nil {
+		t.Errorf("Shutdown error: %v", err)
+	}
+}
+
+func TestOtelAdapterAllInstruments(t *testing.T) {
+	if os.Getenv("SKIP_OTEL_INTEGRATION") != "" {
+		t.Skip("skipping OTel integration test because SKIP_OTEL_INTEGRATION is set")
+	}
+
+	dialTimeout := 300 * time.Millisecond
+	conn, err := net.DialTimeout("tcp", otelGRPCEndpoint, dialTimeout)
+	if err != nil {
+		t.Skipf("skipping OTel integration test; collector %s not reachable: %v", otelGRPCEndpoint, err)
+	}
+	_ = conn.Close()
+
+	b, err := observability.New(observability.MetricsConfig{
+		Enabled: true, Backend: "otel",
+		OTel: observability.OTelConfig{Protocol: "grpc", Endpoint: otelGRPCEndpoint, Insecure: true, ExportInterval: 100 * time.Millisecond, Timeout: 2 * time.Second},
+	})
+	if err != nil {
+		t.Fatalf("failed to create otel backend: %v", err)
+	}
+	ctx := context.Background()
+	labels := observability.Labels{"k": "v"}
+
+	c, err := b.NewCounter("otel_adapter_counter_total", "desc", "k")
+	if err != nil {
+		t.Fatalf("NewCounter error: %v", err)
+	}
+	u, err := b.NewUpDownCounter("otel_adapter_updown", "desc", "k")
+	if err != nil {
+		t.Fatalf("NewUpDownCounter error: %v", err)
+	}
+	ig, err := b.NewInt64Gauge("otel_adapter_int_gauge", "desc", "k")
+	if err != nil {
+		t.Fatalf("NewInt64Gauge error: %v", err)
+	}
+	fg, err := b.NewFloat64Gauge("otel_adapter_float_gauge", "desc", "k")
+	if err != nil {
+		t.Fatalf("NewFloat64Gauge error: %v", err)
+	}
+	h, err := b.NewHistogram("otel_adapter_histogram", "desc", []float64{0.1, 1.0}, "k")
+	if err != nil {
+		t.Fatalf("NewHistogram error: %v", err)
+	}
+
+	c.Add(ctx, 1, labels)
+	u.Add(ctx, 2, labels)
+	ig.Record(ctx, 99, labels)
+	fg.Record(ctx, 1.23, labels)
+	h.Record(ctx, 0.5, labels)
+
+	if b.HTTPHandler() != nil {
+		t.Error("OTel adapter HTTPHandler should be nil")
+	}
+
+	shutdownCtx, cancel := context.WithTimeout(ctx, 2*time.Second)
+	defer cancel()
+	b.Shutdown(shutdownCtx) //nolint:errcheck
+}

internal/observability/noop.go

@@ -0,0 +1,64 @@
+package observability
+
+import (
+	"context"
+	"net/http"
+)
+
+// NoopBackend is a Backend that discards all observations.
+// It is used when metrics are disabled (Enabled=false or Backend="none").
+// All methods are safe to call concurrently.
+type NoopBackend struct{}
+
+// Compile-time interface check.
+var _ Backend = (*NoopBackend)(nil)
+
+func (n *NoopBackend) NewCounter(_ string, _ string, _ ...string) (Counter, error) {
+	return noopCounter{}, nil
+}
+
+func (n *NoopBackend) NewUpDownCounter(_ string, _ string, _ ...string) (UpDownCounter, error) {
+	return noopUpDownCounter{}, nil
+}
+
+func (n *NoopBackend) NewInt64Gauge(_ string, _ string, _ ...string) (Int64Gauge, error) {
+	return noopInt64Gauge{}, nil
+}
+
+func (n *NoopBackend) NewFloat64Gauge(_ string, _ string, _ ...string) (Float64Gauge, error) {
+	return noopFloat64Gauge{}, nil
+}
+
+func (n *NoopBackend) NewHistogram(_ string, _ string, _ []float64, _ ...string) (Histogram, error) {
+	return noopHistogram{}, nil
+}
+
+func (n *NoopBackend) HTTPHandler() http.Handler {
+	return nil
+}
+
+func (n *NoopBackend) Shutdown(_ context.Context) error {
+	return nil
+}
+
+// --- noop instrument types ---
+
+type noopCounter struct{}
+
+func (noopCounter) Add(_ context.Context, _ int64, _ Labels) { /* intentionally no-op */ }
+
+type noopUpDownCounter struct{}
+
+func (noopUpDownCounter) Add(_ context.Context, _ int64, _ Labels) { /* intentionally no-op */ }
+
+type noopInt64Gauge struct{}
+
+func (noopInt64Gauge) Record(_ context.Context, _ int64, _ Labels) { /* intentionally no-op */ }
+
+type noopFloat64Gauge struct{}
+
+func (noopFloat64Gauge) Record(_ context.Context, _ float64, _ Labels) { /* intentionally no-op */ }
+
+type noopHistogram struct{}
+
+func (noopHistogram) Record(_ context.Context, _ float64, _ Labels) { /* intentionally no-op */ }

internal/observability/noop_test.go

@@ -0,0 +1,102 @@
+package observability_test
+
+import (
+	"context"
+	"testing"
+
+	"github.com/fosrl/gerbil/internal/observability"
+)
+
+func TestNoopBackendAllInstruments(t *testing.T) {
+	n := &observability.NoopBackend{}
+
+	ctx := context.Background()
+	labels := observability.Labels{"k": "v"}
+
+	t.Run("Counter", func(t *testing.T) {
+		c, _ := n.NewCounter("test_counter", "desc")
+		c.Add(ctx, 1, labels)
+		c.Add(ctx, 0, nil)
+	})
+
+	t.Run("UpDownCounter", func(t *testing.T) {
+		u, _ := n.NewUpDownCounter("test_updown", "desc")
+		u.Add(ctx, 1, labels)
+		u.Add(ctx, -1, nil)
+	})
+
+	t.Run("Int64Gauge", func(t *testing.T) {
+		g, _ := n.NewInt64Gauge("test_int64gauge", "desc")
+		g.Record(ctx, 42, labels)
+		g.Record(ctx, 0, nil)
+	})
+
+	t.Run("Float64Gauge", func(t *testing.T) {
+		g, _ := n.NewFloat64Gauge("test_float64gauge", "desc")
+		g.Record(ctx, 3.14, labels)
+		g.Record(ctx, 0, nil)
+	})
+
+	t.Run("Histogram", func(t *testing.T) {
+		h, _ := n.NewHistogram("test_histogram", "desc", []float64{1, 5, 10})
+		h.Record(ctx, 2.5, labels)
+		h.Record(ctx, 0, nil)
+	})
+
+	t.Run("HTTPHandler", func(t *testing.T) {
+		if n.HTTPHandler() != nil {
+			t.Error("noop HTTPHandler should be nil")
+		}
+	})
+
+	t.Run("Shutdown", func(t *testing.T) {
+		if err := n.Shutdown(ctx); err != nil {
+			t.Errorf("noop Shutdown should not error: %v", err)
+		}
+	})
+}
+
+func TestNoopBackendLabelNames(t *testing.T) {
+	// Verify that label names passed at creation time are accepted without panic.
+	n := &observability.NoopBackend{}
+
+	assertNoPanic := func(t *testing.T, constructor string, fn func()) {
+		t.Helper()
+		defer func() {
+			if r := recover(); r != nil {
+				t.Fatalf("%s panicked: %v", constructor, r)
+			}
+		}()
+		fn()
+	}
+
+	t.Run("NewCounter", func(t *testing.T) {
+		assertNoPanic(t, "NewCounter", func() {
+			_, _ = n.NewCounter("c", "d", "label1", "label2")
+		})
+	})
+
+	t.Run("NewUpDownCounter", func(t *testing.T) {
+		assertNoPanic(t, "NewUpDownCounter", func() {
+			_, _ = n.NewUpDownCounter("u", "d", "l1")
+		})
+	})
+
+	t.Run("NewInt64Gauge", func(t *testing.T) {
+		assertNoPanic(t, "NewInt64Gauge", func() {
+			_, _ = n.NewInt64Gauge("g1", "d", "l1", "l2", "l3")
+		})
+	})
+
+	t.Run("NewFloat64Gauge", func(t *testing.T) {
+		assertNoPanic(t, "NewFloat64Gauge", func() {
+			_, _ = n.NewFloat64Gauge("g2", "d")
+		})
+	})
+
+	t.Run("NewHistogram", func(t *testing.T) {
+		assertNoPanic(t, "NewHistogram", func() {
+			_, _ = n.NewHistogram("h", "d", []float64{0.1, 1.0}, "l1")
+		})
+	})
+}

internal/observability/otel/backend.go

@@ -0,0 +1,309 @@
+// Package otel implements the OpenTelemetry metrics backend for Gerbil.
+//
+// Metrics are exported via OTLP (gRPC or HTTP) to an external collector.
+// No Prometheus /metrics endpoint is exposed in this mode.
+// Future OTel tracing and logging can be added alongside this package
+// without touching the Prometheus-native path.
+package otel
+
+import (
+	"context"
+	"fmt"
+	"log"
+	"net/http"
+	"regexp"
+	"strings"
+	"time"
+
+	"go.opentelemetry.io/otel/attribute"
+	"go.opentelemetry.io/otel/metric"
+	sdkmetric "go.opentelemetry.io/otel/sdk/metric"
+)
+
+var metricLabelNameRE = regexp.MustCompile(`^[a-zA-Z_][a-zA-Z0-9_]*$`)
+
+// Config holds OTel backend configuration.
+type Config struct {
+	// Protocol is "grpc" (default) or "http".
+	Protocol string
+
+	// Endpoint is the OTLP collector address.
+	Endpoint string
+
+	// Insecure disables TLS.
+	Insecure bool
+
+	// ExportInterval is the period between pushes to the collector.
+	ExportInterval time.Duration
+
+	// Timeout bounds exporter construction calls.
+	Timeout time.Duration
+
+	ServiceName           string
+	ServiceVersion        string
+	DeploymentEnvironment string
+}
+
+// Backend is the OTel metrics backend.
+type Backend struct {
+	cfg      Config
+	provider *sdkmetric.MeterProvider
+	meter    metric.Meter
+}
+
+// New creates and initialises an OTel backend.
+//
+// cfg.Protocol must be "grpc" (default) or "http".
+// cfg.Endpoint is the OTLP collector address (e.g. "localhost:4317").
+// cfg.ExportInterval sets the push period (defaults to 60 s if ≤ 0).
+// cfg.Insecure disables TLS on the OTLP connection.
+//
+// Connection to the collector is established lazily; New only validates cfg
+// and creates the SDK components. It returns an error only if the OTel resource
+// or exporter cannot be constructed.
+func New(cfg Config) (*Backend, error) {
+	if cfg.Protocol == "" {
+		cfg.Protocol = "grpc"
+	}
+	if strings.TrimSpace(cfg.Endpoint) == "" {
+		return nil, fmt.Errorf("otel backend: empty cfg.Endpoint")
+	}
+	if cfg.ExportInterval <= 0 {
+		cfg.ExportInterval = 60 * time.Second
+	}
+	if cfg.Timeout <= 0 {
+		cfg.Timeout = 10 * time.Second
+	}
+	if cfg.ServiceName == "" {
+		cfg.ServiceName = "gerbil"
+	}
+
+	res, err := newResource(cfg.ServiceName, cfg.ServiceVersion, cfg.DeploymentEnvironment)
+	if err != nil {
+		return nil, fmt.Errorf("otel backend: build resource: %w", err)
+	}
+
+	exp, err := newExporter(context.Background(), cfg)
+	if err != nil {
+		return nil, fmt.Errorf("otel backend: create exporter: %w", err)
+	}
+
+	reader := sdkmetric.NewPeriodicReader(exp,
+		sdkmetric.WithInterval(cfg.ExportInterval),
+	)
+
+	provider := sdkmetric.NewMeterProvider(
+		sdkmetric.WithResource(res),
+		sdkmetric.WithReader(reader),
+	)
+
+	meter := provider.Meter("github.com/fosrl/gerbil")
+
+	return &Backend{cfg: cfg, provider: provider, meter: meter}, nil
+}
+
+// HTTPHandler returns nil – the OTel backend does not expose an HTTP endpoint.
+func (b *Backend) HTTPHandler() http.Handler {
+	_ = b
+	return nil
+}
+
+// Shutdown flushes pending metrics and shuts down the MeterProvider.
+func (b *Backend) Shutdown(ctx context.Context) error {
+	return b.provider.Shutdown(ctx)
+}
+
+// NewCounter creates an OTel Int64Counter.
+func (b *Backend) NewCounter(name, desc string, labelNames ...string) (*Counter, error) {
+	normalizedLabelNames, err := validateLabelNames(labelNames)
+	if err != nil {
+		return nil, fmt.Errorf("otel: create counter %q: %w", name, err)
+	}
+	c, err := b.meter.Int64Counter(name, metric.WithDescription(desc))
+	if err != nil {
+		return nil, fmt.Errorf("otel: create counter %q: %w", name, err)
+	}
+	return &Counter{c: c, labelNames: normalizedLabelNames}, nil
+}
+
+// NewUpDownCounter creates an OTel Int64UpDownCounter.
+func (b *Backend) NewUpDownCounter(name, desc string, labelNames ...string) (*UpDownCounter, error) {
+	normalizedLabelNames, err := validateLabelNames(labelNames)
+	if err != nil {
+		return nil, fmt.Errorf("otel: create up-down counter %q: %w", name, err)
+	}
+	c, err := b.meter.Int64UpDownCounter(name, metric.WithDescription(desc))
+	if err != nil {
+		return nil, fmt.Errorf("otel: create up-down counter %q: %w", name, err)
+	}
+	return &UpDownCounter{c: c, labelNames: normalizedLabelNames}, nil
+}
+
+// NewInt64Gauge creates an OTel Int64Gauge.
+func (b *Backend) NewInt64Gauge(name, desc string, labelNames ...string) (*Int64Gauge, error) {
+	normalizedLabelNames, err := validateLabelNames(labelNames)
+	if err != nil {
+		return nil, fmt.Errorf("otel: create int64 gauge %q: %w", name, err)
+	}
+	g, err := b.meter.Int64Gauge(name, metric.WithDescription(desc))
+	if err != nil {
+		return nil, fmt.Errorf("otel: create int64 gauge %q: %w", name, err)
+	}
+	return &Int64Gauge{g: g, labelNames: normalizedLabelNames}, nil
+}
+
+// NewFloat64Gauge creates an OTel Float64Gauge.
+func (b *Backend) NewFloat64Gauge(name, desc string, labelNames ...string) (*Float64Gauge, error) {
+	normalizedLabelNames, err := validateLabelNames(labelNames)
+	if err != nil {
+		return nil, fmt.Errorf("otel: create float64 gauge %q: %w", name, err)
+	}
+	g, err := b.meter.Float64Gauge(name, metric.WithDescription(desc))
+	if err != nil {
+		return nil, fmt.Errorf("otel: create float64 gauge %q: %w", name, err)
+	}
+	return &Float64Gauge{g: g, labelNames: normalizedLabelNames}, nil
+}
+
+// NewHistogram creates an OTel Float64Histogram with explicit bucket boundaries.
+func (b *Backend) NewHistogram(name, desc string, buckets []float64, labelNames ...string) (*Histogram, error) {
+	normalizedLabelNames, err := validateLabelNames(labelNames)
+	if err != nil {
+		return nil, fmt.Errorf("otel: create histogram %q: %w", name, err)
+	}
+	h, err := b.meter.Float64Histogram(name,
+		metric.WithDescription(desc),
+		metric.WithExplicitBucketBoundaries(buckets...),
+	)
+	if err != nil {
+		return nil, fmt.Errorf("otel: create histogram %q: %w", name, err)
+	}
+	return &Histogram{h: h, labelNames: normalizedLabelNames}, nil
+}
+
+func validateLabelNames(labelNames []string) ([]string, error) {
+	if len(labelNames) == 0 {
+		return nil, nil
+	}
+
+	normalized := make([]string, len(labelNames))
+	seen := make(map[string]struct{}, len(labelNames))
+	for i, name := range labelNames {
+		if !metricLabelNameRE.MatchString(name) {
+			return nil, fmt.Errorf("invalid label name %q", name)
+		}
+		if _, exists := seen[name]; exists {
+			return nil, fmt.Errorf("duplicate label name %q", name)
+		}
+		seen[name] = struct{}{}
+		normalized[i] = name
+	}
+
+	return normalized, nil
+}
+
+func labelsToAttrs(labelNames []string, labels map[string]string) []attribute.KeyValue {
+	if len(labelNames) == 0 {
+		if len(labels) > 0 {
+			log.Printf("WARN: dropping otel metric sample due to unexpected labels: got=%v expected=none", labels)
+			return nil
+		}
+		return []attribute.KeyValue{}
+	}
+
+	attrs := make([]attribute.KeyValue, 0, len(labelNames))
+	for _, labelName := range labelNames {
+		attrs = append(attrs, attribute.String(labelName, labels[labelName]))
+	}
+
+	for got := range labels {
+		found := false
+		for _, expected := range labelNames {
+			if got == expected {
+				found = true
+				break
+			}
+		}
+		if !found {
+			log.Printf("WARN: dropping otel metric sample due to unexpected label key %q (expected=%v)", got, labelNames)
+			return nil
+		}
+	}
+
+	return attrs
+}
+
+// Counter wraps an OTel Int64Counter.
+type Counter struct {
+	c          metric.Int64Counter
+	labelNames []string
+}
+
+// Add increments the counter by value.
+func (c *Counter) Add(ctx context.Context, value int64, labels map[string]string) {
+	attrs := labelsToAttrs(c.labelNames, labels)
+	if attrs == nil {
+		return
+	}
+	c.c.Add(ctx, value, metric.WithAttributes(attrs...))
+}
+
+// UpDownCounter wraps an OTel Int64UpDownCounter.
+type UpDownCounter struct {
+	c          metric.Int64UpDownCounter
+	labelNames []string
+}
+
+// Add adjusts the up-down counter by value.
+func (u *UpDownCounter) Add(ctx context.Context, value int64, labels map[string]string) {
+	attrs := labelsToAttrs(u.labelNames, labels)
+	if attrs == nil {
+		return
+	}
+	u.c.Add(ctx, value, metric.WithAttributes(attrs...))
+}
+
+// Int64Gauge wraps an OTel Int64Gauge.
+type Int64Gauge struct {
+	g          metric.Int64Gauge
+	labelNames []string
+}
+
+// Record sets the gauge to value.
+func (g *Int64Gauge) Record(ctx context.Context, value int64, labels map[string]string) {
+	attrs := labelsToAttrs(g.labelNames, labels)
+	if attrs == nil {
+		return
+	}
+	g.g.Record(ctx, value, metric.WithAttributes(attrs...))
+}
+
+// Float64Gauge wraps an OTel Float64Gauge.
+type Float64Gauge struct {
+	g          metric.Float64Gauge
+	labelNames []string
+}
+
+// Record sets the gauge to value.
+func (g *Float64Gauge) Record(ctx context.Context, value float64, labels map[string]string) {
+	attrs := labelsToAttrs(g.labelNames, labels)
+	if attrs == nil {
+		return
+	}
+	g.g.Record(ctx, value, metric.WithAttributes(attrs...))
+}
+
+// Histogram wraps an OTel Float64Histogram.
+type Histogram struct {
+	h          metric.Float64Histogram
+	labelNames []string
+}
+
+// Record observes value in the histogram.
+func (h *Histogram) Record(ctx context.Context, value float64, labels map[string]string) {
+	attrs := labelsToAttrs(h.labelNames, labels)
+	if attrs == nil {
+		return
+	}
+	h.h.Record(ctx, value, metric.WithAttributes(attrs...))
+}

internal/observability/otel/backend_test.go

@@ -0,0 +1,175 @@
+package otel_test
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	obsotel "github.com/fosrl/gerbil/internal/observability/otel"
+)
+
+const (
+	defaultGRPCEndpoint = "localhost:4317"
+	defaultServiceName  = "gerbil-test"
+)
+
+func newInMemoryBackend(t *testing.T) *obsotel.Backend {
+	t.Helper()
+	// Use a very short export interval; an in-process collector (noop exporter)
+	// is used by pointing to a non-existent endpoint with insecure mode.
+	// The backend itself should initialise without error since connection is lazy.
+	b, err := obsotel.New(obsotel.Config{
+		Protocol:       "grpc",
+		Endpoint:       defaultGRPCEndpoint,
+		Insecure:       true,
+		ExportInterval: 100 * time.Millisecond,
+		ServiceName:    defaultServiceName,
+		ServiceVersion: "0.0.1",
+	})
+	if err != nil {
+		t.Fatalf("failed to create otel backend: %v", err)
+	}
+	return b
+}
+
+func TestOtelBackendHTTPHandlerIsNil(t *testing.T) {
+	b := newInMemoryBackend(t)
+	defer b.Shutdown(context.Background()) //nolint:errcheck
+	if b.HTTPHandler() != nil {
+		t.Error("OTel backend HTTPHandler should return nil")
+	}
+}
+
+func TestOtelBackendShutdown(t *testing.T) {
+	b := newInMemoryBackend(t)
+	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Second)
+	defer cancel()
+	if err := b.Shutdown(ctx); err != nil {
+		// Shutdown with unreachable collector may fail to flush; that's acceptable.
+		// What matters is that Shutdown does not panic.
+		t.Logf("Shutdown returned (expected with no collector): %v", err)
+	}
+}
+
+func TestOtelBackendCounter(t *testing.T) {
+	b := newInMemoryBackend(t)
+	defer b.Shutdown(context.Background()) //nolint:errcheck
+
+	c, err := b.NewCounter("gerbil_test_counter_total", "test counter", "result")
+	if err != nil {
+		t.Fatalf("NewCounter returned error: %v", err)
+	}
+	// Should not panic
+	c.Add(context.Background(), 1, map[string]string{"result": "ok"})
+	c.Add(context.Background(), 5, nil)
+}
+
+func TestOtelBackendUpDownCounter(t *testing.T) {
+	b := newInMemoryBackend(t)
+	defer b.Shutdown(context.Background()) //nolint:errcheck
+
+	u, err := b.NewUpDownCounter("gerbil_test_updown", "test updown", "state")
+	if err != nil {
+		t.Fatalf("NewUpDownCounter returned error: %v", err)
+	}
+	u.Add(context.Background(), 3, map[string]string{"state": "active"})
+	u.Add(context.Background(), -1, map[string]string{"state": "active"})
+}
+
+func TestOtelBackendInt64Gauge(t *testing.T) {
+	b := newInMemoryBackend(t)
+	defer b.Shutdown(context.Background()) //nolint:errcheck
+
+	g, err := b.NewInt64Gauge("gerbil_test_int_gauge", "test gauge")
+	if err != nil {
+		t.Fatalf("NewInt64Gauge returned error: %v", err)
+	}
+	g.Record(context.Background(), 42, nil)
+}
+
+func TestOtelBackendFloat64Gauge(t *testing.T) {
+	b := newInMemoryBackend(t)
+	defer b.Shutdown(context.Background()) //nolint:errcheck
+
+	g, err := b.NewFloat64Gauge("gerbil_test_float_gauge", "test float gauge")
+	if err != nil {
+		t.Fatalf("NewFloat64Gauge returned error: %v", err)
+	}
+	g.Record(context.Background(), 3.14, nil)
+}
+
+func TestOtelBackendHistogram(t *testing.T) {
+	b := newInMemoryBackend(t)
+	defer b.Shutdown(context.Background()) //nolint:errcheck
+
+	h, err := b.NewHistogram("gerbil_test_duration_seconds", "test histogram",
+		[]float64{0.1, 0.5, 1.0}, "method")
+	if err != nil {
+		t.Fatalf("NewHistogram returned error: %v", err)
+	}
+	h.Record(context.Background(), 0.3, map[string]string{"method": "GET"})
+}
+
+func TestOtelBackendHTTPProtocol(t *testing.T) {
+	b, err := obsotel.New(obsotel.Config{
+		Protocol:       "http",
+		Endpoint:       "localhost:4318",
+		Insecure:       true,
+		ExportInterval: 100 * time.Millisecond,
+		ServiceName:    defaultServiceName,
+	})
+	if err != nil {
+		t.Fatalf("failed to create otel http backend: %v", err)
+	}
+	defer b.Shutdown(context.Background()) //nolint:errcheck
+
+	if b.HTTPHandler() != nil {
+		t.Error("OTel HTTP backend should not expose a /metrics endpoint")
+	}
+}
+
+func TestOtelBackendInvalidProtocol(t *testing.T) {
+	_, err := obsotel.New(obsotel.Config{
+		Protocol:       "tcp",
+		Endpoint:       defaultGRPCEndpoint,
+		ExportInterval: 10 * time.Second,
+	})
+	if err == nil {
+		t.Error("expected error for invalid protocol")
+	}
+}
+
+func TestOtelBackendDeploymentEnvironment(t *testing.T) {
+	b, err := obsotel.New(obsotel.Config{
+		Protocol:              "grpc",
+		Endpoint:              defaultGRPCEndpoint,
+		Insecure:              true,
+		ExportInterval:        100 * time.Millisecond,
+		ServiceName:           defaultServiceName,
+		ServiceVersion:        "1.2.3",
+		DeploymentEnvironment: "staging",
+	})
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	defer b.Shutdown(context.Background()) //nolint:errcheck
+}
+
+func TestOtelBackendRejectsInvalidLabelNames(t *testing.T) {
+	b := newInMemoryBackend(t)
+	defer b.Shutdown(context.Background()) //nolint:errcheck
+
+	t.Run("duplicate labels", func(t *testing.T) {
+		_, err := b.NewCounter("gerbil_test_invalid_labels_total", "test counter", "result", "result")
+		if err == nil {
+			t.Fatal("expected error for duplicate label names")
+		}
+	})
+
+	t.Run("invalid label name", func(t *testing.T) {
+		_, err := b.NewHistogram("gerbil_test_invalid_histogram", "test histogram", []float64{0.1, 1.0}, "status-code")
+		if err == nil {
+			t.Fatal("expected error for invalid label name")
+		}
+	})
+}

internal/observability/otel/exporter.go

@@ -0,0 +1,68 @@
+package otel
+
+import (
+	"context"
+	"fmt"
+	"net/url"
+	"strings"
+
+	"go.opentelemetry.io/otel/exporters/otlp/otlpmetric/otlpmetricgrpc"
+	"go.opentelemetry.io/otel/exporters/otlp/otlpmetric/otlpmetrichttp"
+	sdkmetric "go.opentelemetry.io/otel/sdk/metric"
+)
+
+// newExporter creates the appropriate OTLP exporter based on cfg.Protocol.
+func newExporter(ctx context.Context, cfg Config) (sdkmetric.Exporter, error) {
+	if strings.TrimSpace(cfg.Endpoint) == "" {
+		return nil, fmt.Errorf("otel: cfg.Endpoint is empty")
+	}
+
+	switch cfg.Protocol {
+	case "grpc", "":
+		return newGRPCExporter(ctx, cfg)
+	case "http":
+		return newHTTPExporter(ctx, cfg)
+	default:
+		return nil, fmt.Errorf("otel: unknown protocol %q (must be \"grpc\" or \"http\")", cfg.Protocol)
+	}
+}
+
+func newGRPCExporter(ctx context.Context, cfg Config) (sdkmetric.Exporter, error) {
+	opts := []otlpmetricgrpc.Option{
+		otlpmetricgrpc.WithEndpoint(cfg.Endpoint),
+	}
+	if cfg.Insecure {
+		opts = append(opts, otlpmetricgrpc.WithInsecure())
+	}
+	exp, err := otlpmetricgrpc.New(ctx, opts...)
+	if err != nil {
+		return nil, fmt.Errorf("otlp grpc exporter: %w", err)
+	}
+	return exp, nil
+}
+
+func newHTTPExporter(ctx context.Context, cfg Config) (sdkmetric.Exporter, error) {
+	endpoint := strings.TrimSpace(cfg.Endpoint)
+
+	opts := make([]otlpmetrichttp.Option, 0, 3)
+	if strings.Contains(endpoint, "://") {
+		parsed, err := url.Parse(endpoint)
+		if err != nil {
+			return nil, fmt.Errorf("otlp http exporter: parse endpoint URL %q: %w", endpoint, err)
+		}
+		opts = append(opts, otlpmetrichttp.WithEndpointURL(parsed.String()))
+	} else {
+		opts = append(opts,
+			otlpmetrichttp.WithEndpoint(endpoint),
+			otlpmetrichttp.WithURLPath("/v1/metrics"),
+		)
+	}
+	if cfg.Insecure {
+		opts = append(opts, otlpmetrichttp.WithInsecure())
+	}
+	exp, err := otlpmetrichttp.New(ctx, opts...)
+	if err != nil {
+		return nil, fmt.Errorf("otlp http exporter: %w", err)
+	}
+	return exp, nil
+}

internal/observability/otel/resource.go

@@ -0,0 +1,25 @@
+package otel
+
+import (
+	"go.opentelemetry.io/otel/attribute"
+	"go.opentelemetry.io/otel/sdk/resource"
+	semconv "go.opentelemetry.io/otel/semconv/v1.18.0"
+)
+
+// newResource builds an OTel resource for the Gerbil service.
+func newResource(serviceName, serviceVersion, deploymentEnv string) (*resource.Resource, error) {
+	attrs := []attribute.KeyValue{
+		semconv.ServiceName(serviceName),
+	}
+	if serviceVersion != "" {
+		attrs = append(attrs, semconv.ServiceVersion(serviceVersion))
+	}
+	if deploymentEnv != "" {
+		attrs = append(attrs, semconv.DeploymentEnvironment(deploymentEnv))
+	}
+
+	return resource.Merge(
+		resource.Default(),
+		resource.NewSchemaless(attrs...),
+	)
+}

internal/observability/prometheus/backend.go

@@ -0,0 +1,310 @@
+// Package prometheus implements the native Prometheus metrics backend for Gerbil.
+//
+// This backend uses the Prometheus Go client directly; it does NOT depend on the
+// OpenTelemetry SDK.  A dedicated Prometheus registry is used so that default
+// Go/process metrics are not unintentionally included unless the caller opts in.
+package prometheus
+
+import (
+	"context"
+	"log"
+	"net/http"
+
+	"github.com/prometheus/client_golang/prometheus"
+	"github.com/prometheus/client_golang/prometheus/collectors"
+	"github.com/prometheus/client_golang/prometheus/promhttp"
+)
+
+// Config holds Prometheus-backend configuration.
+type Config struct {
+	// Path is the HTTP endpoint path (e.g. "/metrics").
+	Path string
+
+	// IncludeGoMetrics controls whether the standard Go runtime and process
+	// collectors are registered on the dedicated registry.
+	// Defaults to true if not explicitly set.
+	IncludeGoMetrics *bool
+}
+
+// Backend is the native Prometheus metrics backend.
+// Metric instruments are created via the New* family of methods and stored
+// in the backend-specific instrument types that implement the observability
+// instrument interfaces.
+type Backend struct {
+	cfg                   Config
+	registry              *prometheus.Registry
+	handler               http.Handler
+	droppedSamplesCounter prometheus.Counter
+}
+
+// New creates and initialises a Prometheus backend.
+//
+// cfg.Path sets the HTTP endpoint path (defaults to "/metrics" if empty).
+// cfg.IncludeGoMetrics controls whether standard Go runtime and process metrics
+// are included; defaults to true when nil.
+//
+// Returns an error if the registry cannot be created.
+func New(cfg Config) (*Backend, error) {
+	if cfg.Path == "" {
+		cfg.Path = "/metrics"
+	}
+
+	registry := prometheus.NewRegistry()
+	droppedSamplesCounter := prometheus.NewCounter(prometheus.CounterOpts{
+		Name: "gerbil_dropped_metric_samples_total",
+		Help: "Total number of metric samples dropped due to invalid labels or unsupported label sets",
+	})
+	registry.MustRegister(droppedSamplesCounter)
+
+	// Include Go and process metrics by default.
+	includeGo := cfg.IncludeGoMetrics == nil || *cfg.IncludeGoMetrics
+	if includeGo {
+		registry.MustRegister(
+			collectors.NewGoCollector(),
+			collectors.NewProcessCollector(collectors.ProcessCollectorOpts{}),
+		)
+	}
+
+	handler := promhttp.HandlerFor(registry, promhttp.HandlerOpts{
+		EnableOpenMetrics: false,
+	})
+
+	return &Backend{cfg: cfg, registry: registry, handler: handler, droppedSamplesCounter: droppedSamplesCounter}, nil
+}
+
+// HTTPHandler returns the Prometheus /metrics HTTP handler.
+func (b *Backend) HTTPHandler() http.Handler {
+	return b.handler
+}
+
+// Shutdown is a no-op for the Prometheus backend.
+// The registry does not maintain background goroutines.
+func (b *Backend) Shutdown(_ context.Context) error {
+	_ = b
+	return nil
+}
+
+// NewCounter creates a Prometheus CounterVec registered on the backend's registry.
+func (b *Backend) NewCounter(name, desc string, labelNames ...string) (*Counter, error) {
+	vec := prometheus.NewCounterVec(prometheus.CounterOpts{
+		Name: name,
+		Help: desc,
+	}, labelNames)
+	if err := b.registry.Register(vec); err != nil {
+		if are, ok := err.(prometheus.AlreadyRegisteredError); ok {
+			existing, ok := are.ExistingCollector.(*prometheus.CounterVec)
+			if !ok {
+				return nil, err
+			}
+			return &Counter{vec: existing, labelNames: append([]string(nil), labelNames...), droppedSamplesCounter: b.droppedSamplesCounter}, nil
+		}
+		return nil, err
+	}
+	return &Counter{vec: vec, labelNames: append([]string(nil), labelNames...), droppedSamplesCounter: b.droppedSamplesCounter}, nil
+}
+
+// NewUpDownCounter creates a Prometheus GaugeVec (Prometheus gauges are
+// bidirectional) registered on the backend's registry.
+func (b *Backend) NewUpDownCounter(name, desc string, labelNames ...string) (*UpDownCounter, error) {
+	vec := prometheus.NewGaugeVec(prometheus.GaugeOpts{
+		Name: name,
+		Help: desc,
+	}, labelNames)
+	if err := b.registry.Register(vec); err != nil {
+		if are, ok := err.(prometheus.AlreadyRegisteredError); ok {
+			existing, ok := are.ExistingCollector.(*prometheus.GaugeVec)
+			if !ok {
+				return nil, err
+			}
+			return &UpDownCounter{vec: existing, labelNames: append([]string(nil), labelNames...), droppedSamplesCounter: b.droppedSamplesCounter}, nil
+		}
+		return nil, err
+	}
+	return &UpDownCounter{vec: vec, labelNames: append([]string(nil), labelNames...), droppedSamplesCounter: b.droppedSamplesCounter}, nil
+}
+
+// NewInt64Gauge creates a Prometheus GaugeVec registered on the backend's registry.
+func (b *Backend) NewInt64Gauge(name, desc string, labelNames ...string) (*Int64Gauge, error) {
+	vec := prometheus.NewGaugeVec(prometheus.GaugeOpts{
+		Name: name,
+		Help: desc,
+	}, labelNames)
+	if err := b.registry.Register(vec); err != nil {
+		if are, ok := err.(prometheus.AlreadyRegisteredError); ok {
+			existing, ok := are.ExistingCollector.(*prometheus.GaugeVec)
+			if !ok {
+				return nil, err
+			}
+			return &Int64Gauge{vec: existing, labelNames: append([]string(nil), labelNames...), droppedSamplesCounter: b.droppedSamplesCounter}, nil
+		}
+		return nil, err
+	}
+	return &Int64Gauge{vec: vec, labelNames: append([]string(nil), labelNames...), droppedSamplesCounter: b.droppedSamplesCounter}, nil
+}
+
+// NewFloat64Gauge creates a Prometheus GaugeVec registered on the backend's registry.
+func (b *Backend) NewFloat64Gauge(name, desc string, labelNames ...string) (*Float64Gauge, error) {
+	vec := prometheus.NewGaugeVec(prometheus.GaugeOpts{
+		Name: name,
+		Help: desc,
+	}, labelNames)
+	if err := b.registry.Register(vec); err != nil {
+		if are, ok := err.(prometheus.AlreadyRegisteredError); ok {
+			existing, ok := are.ExistingCollector.(*prometheus.GaugeVec)
+			if !ok {
+				return nil, err
+			}
+			return &Float64Gauge{vec: existing, labelNames: append([]string(nil), labelNames...), droppedSamplesCounter: b.droppedSamplesCounter}, nil
+		}
+		return nil, err
+	}
+	return &Float64Gauge{vec: vec, labelNames: append([]string(nil), labelNames...), droppedSamplesCounter: b.droppedSamplesCounter}, nil
+}
+
+// NewHistogram creates a Prometheus HistogramVec registered on the backend's registry.
+func (b *Backend) NewHistogram(name, desc string, buckets []float64, labelNames ...string) (*Histogram, error) {
+	vec := prometheus.NewHistogramVec(prometheus.HistogramOpts{
+		Name:    name,
+		Help:    desc,
+		Buckets: buckets,
+	}, labelNames)
+	if err := b.registry.Register(vec); err != nil {
+		if are, ok := err.(prometheus.AlreadyRegisteredError); ok {
+			existing, ok := are.ExistingCollector.(*prometheus.HistogramVec)
+			if !ok {
+				return nil, err
+			}
+			return &Histogram{vec: existing, labelNames: append([]string(nil), labelNames...), droppedSamplesCounter: b.droppedSamplesCounter}, nil
+		}
+		return nil, err
+	}
+	return &Histogram{vec: vec, labelNames: append([]string(nil), labelNames...), droppedSamplesCounter: b.droppedSamplesCounter}, nil
+}
+
+// Counter is a native Prometheus counter instrument.
+type Counter struct {
+	vec                   *prometheus.CounterVec
+	labelNames            []string
+	droppedSamplesCounter prometheus.Counter
+}
+
+// Add increments the counter by value for the given labels.
+//
+// value must be non-negative. Negative values are ignored.
+func (c *Counter) Add(_ context.Context, value int64, labels map[string]string) {
+	if value < 0 {
+		log.Printf("WARN: counter add called with negative value=%d labels=%v expected_labels=%v", value, labels, c.labelNames)
+		return
+	}
+	normalized, ok := normalizeLabels(c.labelNames, labels, c.droppedSamplesCounter)
+	if !ok {
+		return
+	}
+	defer guardMetricPanic("counter", c.labelNames, labels)
+	c.vec.With(normalized).Add(float64(value))
+}
+
+// UpDownCounter is a native Prometheus gauge used as a bidirectional counter.
+type UpDownCounter struct {
+	vec                   *prometheus.GaugeVec
+	labelNames            []string
+	droppedSamplesCounter prometheus.Counter
+}
+
+// Add adjusts the gauge by value for the given labels.
+func (u *UpDownCounter) Add(_ context.Context, value int64, labels map[string]string) {
+	normalized, ok := normalizeLabels(u.labelNames, labels, u.droppedSamplesCounter)
+	if !ok {
+		return
+	}
+	defer guardMetricPanic("updown", u.labelNames, labels)
+	u.vec.With(normalized).Add(float64(value))
+}
+
+// Int64Gauge is a native Prometheus gauge recording integer snapshot values.
+type Int64Gauge struct {
+	vec                   *prometheus.GaugeVec
+	labelNames            []string
+	droppedSamplesCounter prometheus.Counter
+}
+
+// Record sets the gauge to value for the given labels.
+func (g *Int64Gauge) Record(_ context.Context, value int64, labels map[string]string) {
+	normalized, ok := normalizeLabels(g.labelNames, labels, g.droppedSamplesCounter)
+	if !ok {
+		return
+	}
+	defer guardMetricPanic("int64-gauge", g.labelNames, labels)
+	g.vec.With(normalized).Set(float64(value))
+}
+
+// Float64Gauge is a native Prometheus gauge recording float snapshot values.
+type Float64Gauge struct {
+	vec                   *prometheus.GaugeVec
+	labelNames            []string
+	droppedSamplesCounter prometheus.Counter
+}
+
+// Record sets the gauge to value for the given labels.
+func (g *Float64Gauge) Record(_ context.Context, value float64, labels map[string]string) {
+	normalized, ok := normalizeLabels(g.labelNames, labels, g.droppedSamplesCounter)
+	if !ok {
+		return
+	}
+	defer guardMetricPanic("float64-gauge", g.labelNames, labels)
+	g.vec.With(normalized).Set(value)
+}
+
+// Histogram is a native Prometheus histogram instrument.
+type Histogram struct {
+	vec                   *prometheus.HistogramVec
+	labelNames            []string
+	droppedSamplesCounter prometheus.Counter
+}
+
+// Record observes value for the given labels.
+func (h *Histogram) Record(_ context.Context, value float64, labels map[string]string) {
+	normalized, ok := normalizeLabels(h.labelNames, labels, h.droppedSamplesCounter)
+	if !ok {
+		return
+	}
+	defer guardMetricPanic("histogram", h.labelNames, labels)
+	h.vec.With(normalized).Observe(value)
+}
+
+func normalizeLabels(labelNames []string, labels map[string]string, droppedSamplesCounter prometheus.Counter) (prometheus.Labels, bool) {
+	if len(labelNames) == 0 {
+		if len(labels) > 0 {
+			if droppedSamplesCounter != nil {
+				droppedSamplesCounter.Inc()
+			}
+			log.Printf("WARN: dropping metric sample due to unexpected labels: got=%v expected=none", labels)
+			return nil, false
+		}
+		return nil, true
+	}
+
+	normalized := make(prometheus.Labels, len(labelNames))
+	for _, name := range labelNames {
+		normalized[name] = ""
+	}
+
+	for k, v := range labels {
+		if _, ok := normalized[k]; !ok {
+			if droppedSamplesCounter != nil {
+				droppedSamplesCounter.Inc()
+			}
+			log.Printf("WARN: dropping metric sample due to unexpected label key %q (expected=%v)", k, labelNames)
+			return nil, false
+		}
+		normalized[k] = v
+	}
+
+	return normalized, true
+}
+
+func guardMetricPanic(kind string, expected []string, labels map[string]string) {
+	if recovered := recover(); recovered != nil {
+		log.Printf("WARN: dropped %s metric sample due to label panic: expected=%v got=%v err=%v", kind, expected, labels, recovered)
+	}
+}

internal/observability/prometheus/backend_test.go

@@ -0,0 +1,231 @@
+package prometheus_test
+
+import (
+	"context"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	obsprom "github.com/fosrl/gerbil/internal/observability/prometheus"
+)
+
+func newTestBackend(t *testing.T) *obsprom.Backend {
+	t.Helper()
+	b, err := obsprom.New(obsprom.Config{Path: "/metrics"})
+	if err != nil {
+		t.Fatalf("failed to create prometheus backend: %v", err)
+	}
+	return b
+}
+
+func TestPrometheusBackendHTTPHandler(t *testing.T) {
+	b := newTestBackend(t)
+	if b.HTTPHandler() == nil {
+		t.Error("HTTPHandler should not be nil")
+	}
+}
+
+func TestPrometheusBackendShutdown(t *testing.T) {
+	b := newTestBackend(t)
+	if err := b.Shutdown(context.Background()); err != nil {
+		t.Errorf("Shutdown returned error: %v", err)
+	}
+}
+
+func TestPrometheusBackendCounter(t *testing.T) {
+	b := newTestBackend(t)
+	c, err := b.NewCounter("test_counter_total", "A test counter", "result")
+	if err != nil {
+		t.Fatalf("NewCounter returned error: %v", err)
+	}
+	c.Add(context.Background(), 3, map[string]string{"result": "ok"})
+
+	body := scrapeMetrics(t, b)
+	assertMetricPresent(t, body, `test_counter_total{result="ok"} 3`)
+}
+
+func TestPrometheusBackendUpDownCounter(t *testing.T) {
+	b := newTestBackend(t)
+	u, err := b.NewUpDownCounter("test_gauge_total", "A test up-down counter", "state")
+	if err != nil {
+		t.Fatalf("NewUpDownCounter returned error: %v", err)
+	}
+	u.Add(context.Background(), 5, map[string]string{"state": "active"})
+	u.Add(context.Background(), -2, map[string]string{"state": "active"})
+
+	body := scrapeMetrics(t, b)
+	assertMetricPresent(t, body, `test_gauge_total{state="active"} 3`)
+}
+
+func TestPrometheusBackendInt64Gauge(t *testing.T) {
+	b := newTestBackend(t)
+	g, err := b.NewInt64Gauge("test_int_gauge", "An integer gauge", "ifname")
+	if err != nil {
+		t.Fatalf("NewInt64Gauge returned error: %v", err)
+	}
+	g.Record(context.Background(), 42, map[string]string{"ifname": "wg0"})
+
+	body := scrapeMetrics(t, b)
+	assertMetricPresent(t, body, `test_int_gauge{ifname="wg0"} 42`)
+}
+
+func TestPrometheusBackendFloat64Gauge(t *testing.T) {
+	b := newTestBackend(t)
+	g, err := b.NewFloat64Gauge("test_float_gauge", "A float gauge", "cert")
+	if err != nil {
+		t.Fatalf("NewFloat64Gauge returned error: %v", err)
+	}
+	g.Record(context.Background(), 7.5, map[string]string{"cert": "example.com"})
+
+	body := scrapeMetrics(t, b)
+	assertMetricPresent(t, body, `test_float_gauge{cert="example.com"} 7.5`)
+}
+
+func TestPrometheusBackendHistogram(t *testing.T) {
+	b := newTestBackend(t)
+	buckets := []float64{0.1, 0.5, 1.0, 5.0}
+	h, err := b.NewHistogram("test_duration_seconds", "A test histogram", buckets, "method")
+	if err != nil {
+		t.Fatalf("NewHistogram returned error: %v", err)
+	}
+	h.Record(context.Background(), 0.3, map[string]string{"method": "GET"})
+
+	body := scrapeMetrics(t, b)
+	if !strings.Contains(body, "test_duration_seconds") {
+		t.Errorf("expected histogram metric in output, body:\n%s", body)
+	}
+}
+
+func TestPrometheusBackendMultipleLabels(t *testing.T) {
+	b := newTestBackend(t)
+	c, err := b.NewCounter("multi_label_total", "Multi-label counter", "method", "route", "status_code")
+	if err != nil {
+		t.Fatalf("NewCounter returned error: %v", err)
+	}
+	c.Add(context.Background(), 1, map[string]string{
+		"method":      "POST",
+		"route":       "/api/peers",
+		"status_code": "200",
+	})
+
+	body := scrapeMetrics(t, b)
+	if !strings.Contains(body, "multi_label_total") {
+		t.Errorf("expected multi_label_total in output, body:\n%s", body)
+	}
+}
+
+func TestPrometheusBackendGoMetrics(t *testing.T) {
+	b := newTestBackend(t)
+	body := scrapeMetrics(t, b)
+	// Default backend includes Go runtime metrics.
+	if !strings.Contains(body, "go_goroutines") {
+		t.Error("expected go_goroutines in default backend output")
+	}
+}
+
+func TestPrometheusBackendNoGoMetrics(t *testing.T) {
+	f := false
+	b, err := obsprom.New(obsprom.Config{IncludeGoMetrics: &f})
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	body := scrapeMetrics(t, b)
+	if strings.Contains(body, "go_goroutines") {
+		t.Error("expected no go_goroutines when IncludeGoMetrics=false")
+	}
+}
+
+func TestPrometheusBackendNilLabels(t *testing.T) {
+	// Adding with nil labels should not panic (treated as empty map).
+	b := newTestBackend(t)
+	c, err := b.NewCounter("nil_labels_total", "counter with no labels")
+	if err != nil {
+		t.Fatalf("NewCounter returned error: %v", err)
+	}
+	// nil labels with no label names declared should be safe
+	c.Add(context.Background(), 1, nil)
+}
+
+func TestPrometheusBackendConcurrentAdd(t *testing.T) {
+	b := newTestBackend(t)
+	c, err := b.NewCounter("concurrent_total", "concurrent counter", "worker")
+	if err != nil {
+		t.Fatalf("NewCounter returned error: %v", err)
+	}
+
+	done := make(chan struct{})
+	for i := 0; i < 10; i++ {
+		go func() {
+			for j := 0; j < 100; j++ {
+				c.Add(context.Background(), 1, map[string]string{"worker": "w"})
+			}
+			done <- struct{}{}
+		}()
+	}
+	for i := 0; i < 10; i++ {
+		<-done
+	}
+
+	body := scrapeMetrics(t, b)
+	assertMetricPresent(t, body, `concurrent_total{worker="w"} 1000`)
+}
+
+func TestPrometheusBackendAlreadyRegisteredCounter(t *testing.T) {
+	b := newTestBackend(t)
+	c1, err := b.NewCounter("dupe_counter_total", "duplicate counter", "result")
+	if err != nil {
+		t.Fatalf("first NewCounter returned error: %v", err)
+	}
+	c2, err := b.NewCounter("dupe_counter_total", "duplicate counter", "result")
+	if err != nil {
+		t.Fatalf("second NewCounter returned error: %v", err)
+	}
+
+	c1.Add(context.Background(), 1, map[string]string{"result": "ok"})
+	c2.Add(context.Background(), 2, map[string]string{"result": "ok"})
+
+	body := scrapeMetrics(t, b)
+	assertMetricPresent(t, body, `dupe_counter_total{result="ok"} 3`)
+}
+
+func TestPrometheusBackendInvalidLabelsNoPanic(t *testing.T) {
+	b := newTestBackend(t)
+	c, err := b.NewCounter("invalid_labels_total", "invalid labels test", "result")
+	if err != nil {
+		t.Fatalf("NewCounter returned error: %v", err)
+	}
+
+	// Extra label key should be dropped and must not panic.
+	c.Add(context.Background(), 5, map[string]string{"result": "ok", "unexpected": "x"})
+
+	body := scrapeMetrics(t, b)
+	if strings.Contains(body, `invalid_labels_total{result="ok"}`) {
+		t.Error("invalid label sample should have been dropped")
+	}
+}
+
+// --- helpers ---
+
+func scrapeMetrics(t *testing.T, b *obsprom.Backend) string {
+	t.Helper()
+	req := httptest.NewRequest(http.MethodGet, "/metrics", http.NoBody)
+	rr := httptest.NewRecorder()
+	b.HTTPHandler().ServeHTTP(rr, req)
+	if rr.Code != http.StatusOK {
+		t.Fatalf("metrics handler returned %d", rr.Code)
+	}
+	body, err := io.ReadAll(rr.Body)
+	if err != nil {
+		t.Fatalf("failed to read response body: %v", err)
+	}
+	return string(body)
+}
+
+func assertMetricPresent(t *testing.T, body, expected string) {
+	t.Helper()
+	if !strings.Contains(body, expected) {
+		t.Errorf("expected %q in metrics output\nbody:\n%s", expected, body)
+	}
+}

proxy/proxy.go

@@ -11,15 +11,58 @@ import (
 	"log"
 	"net"
 	"net/http"
+	"os"
+	"strconv"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"time"
 
+	"github.com/fosrl/gerbil/internal/metrics"
 	"github.com/fosrl/gerbil/logger"
-	"github.com/fosrl/gerbil/proxyproto"
 	"github.com/patrickmn/go-cache"
 )
 
+// defaultMaxSNIConnections caps the number of concurrent client connections
+// the SNI proxy will accept. Without a cap, a burst of connections (a
+// scanner sweep, a client reconnect storm) spawns unbounded goroutines each
+// holding copy buffers and sockets, exhausting container memory faster than
+// GC/backpressure can catch up.
+// Sized conservatively since each connection now holds up to two pooled
+// 32KB copy buffers once the buffer pool is actually honored (see
+// bufferedReader/bufferedWriter below). Overridable via
+// GERBIL_MAX_SNI_CONNECTIONS.
+const defaultMaxSNIConnections = 4096
+
+var maxSNIConnections = loadMaxSNIConnections()
+
+func loadMaxSNIConnections() int64 {
+	if v := os.Getenv("GERBIL_MAX_SNI_CONNECTIONS"); v != "" {
+		if n, err := strconv.ParseInt(v, 10, 64); err == nil && n > 0 {
+			return n
+		}
+	}
+	return defaultMaxSNIConnections
+}
+
+// defaultMaxSNIConnectionsPerIP caps concurrent connections from a single
+// source IP, independent of the global maxSNIConnections budget. Without
+// this, one noisy/misbehaving client (or a single scanning host) can
+// consume the entire global budget and lock out every other customer
+// sharing this proxy. Overridable via GERBIL_MAX_SNI_CONNECTIONS_PER_IP.
+const defaultMaxSNIConnectionsPerIP = 256
+
+var maxSNIConnectionsPerIP = loadMaxSNIConnectionsPerIP()
+
+func loadMaxSNIConnectionsPerIP() int64 {
+	if v := os.Getenv("GERBIL_MAX_SNI_CONNECTIONS_PER_IP"); v != "" {
+		if n, err := strconv.ParseInt(v, 10, 64); err == nil && n > 0 {
+			return n
+		}
+	}
+	return defaultMaxSNIConnectionsPerIP
+}
+
 // RouteRecord represents a routing configuration
 type RouteRecord struct {
 	Hostname   string
@@ -32,6 +75,16 @@ type RouteAPIResponse struct {
 	Endpoints []string `json:"endpoints"`
 }
 
+// ProxyProtocolInfo holds information parsed from incoming PROXY protocol header
+type ProxyProtocolInfo struct {
+	Protocol     string // TCP4 or TCP6
+	SrcIP        string
+	DestIP       string
+	SrcPort      int
+	DestPort     int
+	OriginalConn net.Conn // The original connection after PROXY protocol parsing
+}
+
 // SNIProxy represents the main proxy server
 type SNIProxy struct {
 	port            int
@@ -59,6 +112,24 @@ type SNIProxy struct {
 
 	// Trusted upstream proxies that can send PROXY protocol
 	trustedUpstreams map[string]struct{}
+
+	// Reusable HTTP client for API requests
+	httpClient *http.Client
+
+	// Buffer pool for connection piping
+	bufferPool *sync.Pool
+
+	// activeConnections tracks concurrent client connections so
+	// acceptConnections can enforce maxSNIConnections.
+	activeConnections atomic.Int64
+
+	// perIPConnections tracks concurrent connections per source IP (map[string]*atomic.Int64)
+	// so acceptConnections can enforce maxSNIConnectionsPerIP and stop one
+	// client from starving the rest. Entries are removed once a given IP's
+	// count returns to zero, so this stays bounded by currently-connected
+	// distinct IPs (itself bounded by maxSNIConnections) rather than growing
+	// with every IP ever seen.
+	perIPConnections sync.Map
 }
 
 type activeTunnel struct {
@@ -79,6 +150,249 @@ func (conn readOnlyConn) SetDeadline(t time.Time) error      { return nil }
 func (conn readOnlyConn) SetReadDeadline(t time.Time) error  { return nil }
 func (conn readOnlyConn) SetWriteDeadline(t time.Time) error { return nil }
 
+// parseProxyProtocolHeader parses a PROXY protocol v1 header from the connection
+func (p *SNIProxy) parseProxyProtocolHeader(conn net.Conn) (*ProxyProtocolInfo, net.Conn, error) {
+	// Check if the connection comes from a trusted upstream
+	remoteHost, _, err := net.SplitHostPort(conn.RemoteAddr().String())
+	if err != nil {
+		return nil, conn, fmt.Errorf("failed to parse remote address: %w", err)
+	}
+
+	// Resolve the remote IP to hostname to check if it's trusted
+	// For simplicity, we'll check the IP directly in trusted upstreams
+	// In production, you might want to do reverse DNS lookup
+	if _, isTrusted := p.trustedUpstreams[remoteHost]; !isTrusted {
+		// Not from trusted upstream, return original connection
+		return nil, conn, nil
+	}
+
+	// Set read timeout for PROXY protocol parsing
+	if err := conn.SetReadDeadline(time.Now().Add(5 * time.Second)); err != nil {
+		return nil, conn, fmt.Errorf("failed to set read deadline: %w", err)
+	}
+
+	// Read the first line (PROXY protocol header)
+	buffer := make([]byte, 512) // PROXY protocol header should be much smaller
+	n, err := conn.Read(buffer)
+	if err != nil {
+		// If we can't read from trusted upstream, treat as regular connection
+		logger.Debug("Could not read from trusted upstream %s, treating as regular connection: %v", remoteHost, err)
+		// Clear read timeout before returning
+		if clearErr := conn.SetReadDeadline(time.Time{}); clearErr != nil {
+			logger.Debug("Failed to clear read deadline: %v", clearErr)
+		}
+		return nil, conn, nil
+	}
+
+	// Find the end of the first line (CRLF)
+	headerEnd := bytes.Index(buffer[:n], []byte("\r\n"))
+	if headerEnd == -1 {
+		// No PROXY protocol header found, treat as regular TLS connection
+		// Return the connection with the buffered data prepended
+		logger.Debug("No PROXY protocol header from trusted upstream %s, treating as regular TLS connection", remoteHost)
+
+		// Clear read timeout
+		if err := conn.SetReadDeadline(time.Time{}); err != nil {
+			logger.Debug("Failed to clear read deadline: %v", err)
+		}
+
+		// Create a reader that includes the buffered data + original connection
+		newReader := io.MultiReader(bytes.NewReader(buffer[:n]), conn)
+		wrappedConn := &proxyProtocolConn{
+			Conn:   conn,
+			reader: newReader,
+		}
+		return nil, wrappedConn, nil
+	}
+
+	headerLine := string(buffer[:headerEnd])
+	remainingData := buffer[headerEnd+2 : n]
+
+	// Parse PROXY protocol line: "PROXY TCP4/TCP6 srcIP destIP srcPort destPort"
+	parts := strings.Fields(headerLine)
+	if len(parts) != 6 || parts[0] != "PROXY" {
+		// Check for PROXY UNKNOWN
+		if len(parts) == 2 && parts[0] == "PROXY" && parts[1] == "UNKNOWN" {
+			// PROXY UNKNOWN - use original connection info
+			return nil, conn, nil
+		}
+		// Invalid PROXY protocol, but might be regular TLS - treat as such
+		logger.Debug("Invalid PROXY protocol from trusted upstream %s, treating as regular TLS connection: %s", remoteHost, headerLine)
+
+		// Clear read timeout
+		if err := conn.SetReadDeadline(time.Time{}); err != nil {
+			logger.Debug("Failed to clear read deadline: %v", err)
+		}
+
+		// Return the connection with all buffered data prepended
+		newReader := io.MultiReader(bytes.NewReader(buffer[:n]), conn)
+		wrappedConn := &proxyProtocolConn{
+			Conn:   conn,
+			reader: newReader,
+		}
+		return nil, wrappedConn, nil
+	}
+
+	protocol := parts[1]
+	srcIP := parts[2]
+	destIP := parts[3]
+	srcPort, err := strconv.Atoi(parts[4])
+	if err != nil {
+		return nil, conn, fmt.Errorf("invalid source port in PROXY header: %s", parts[4])
+	}
+	destPort, err := strconv.Atoi(parts[5])
+	if err != nil {
+		return nil, conn, fmt.Errorf("invalid destination port in PROXY header: %s", parts[5])
+	}
+
+	// Create a new reader that includes remaining data + original connection
+	var newReader io.Reader
+	if len(remainingData) > 0 {
+		newReader = io.MultiReader(bytes.NewReader(remainingData), conn)
+	} else {
+		newReader = conn
+	}
+
+	// Create a wrapper connection that reads from the combined reader
+	wrappedConn := &proxyProtocolConn{
+		Conn:   conn,
+		reader: newReader,
+	}
+
+	proxyInfo := &ProxyProtocolInfo{
+		Protocol:     protocol,
+		SrcIP:        srcIP,
+		DestIP:       destIP,
+		SrcPort:      srcPort,
+		DestPort:     destPort,
+		OriginalConn: wrappedConn,
+	}
+
+	// Clear read timeout
+	if err := conn.SetReadDeadline(time.Time{}); err != nil {
+		return nil, conn, fmt.Errorf("failed to clear read deadline: %w", err)
+	}
+
+	return proxyInfo, wrappedConn, nil
+}
+
+// proxyProtocolConn wraps a connection to read from a custom reader
+type proxyProtocolConn struct {
+	net.Conn
+	reader io.Reader
+}
+
+func (c *proxyProtocolConn) Read(b []byte) (int, error) {
+	return c.reader.Read(b)
+}
+
+// buildProxyProtocolHeaderFromInfo creates a PROXY protocol v1 header using ProxyProtocolInfo
+func (p *SNIProxy) buildProxyProtocolHeaderFromInfo(proxyInfo *ProxyProtocolInfo, targetAddr net.Addr) string {
+	targetTCP, ok := targetAddr.(*net.TCPAddr)
+	if !ok {
+		// Fallback for unknown address types
+		return "PROXY UNKNOWN\r\n"
+	}
+
+	// Use the original client information from the PROXY protocol
+	var targetIP string
+	var protocol string
+
+	// Parse source IP to determine protocol family
+	srcIP := net.ParseIP(proxyInfo.SrcIP)
+	if srcIP == nil {
+		return "PROXY UNKNOWN\r\n"
+	}
+
+	if srcIP.To4() != nil {
+		// Source is IPv4, use TCP4 protocol
+		protocol = "TCP4"
+		if targetTCP.IP.To4() != nil {
+			// Target is also IPv4, use as-is
+			targetIP = targetTCP.IP.String()
+		} else {
+			// Target is IPv6, but we need IPv4 for consistent protocol family
+			if targetTCP.IP.IsLoopback() {
+				targetIP = "127.0.0.1"
+			} else {
+				targetIP = "127.0.0.1" // Safe fallback
+			}
+		}
+	} else {
+		// Source is IPv6, use TCP6 protocol
+		protocol = "TCP6"
+		if targetTCP.IP.To4() != nil {
+			// Target is IPv4, convert to IPv6 representation
+			targetIP = "::ffff:" + targetTCP.IP.String()
+		} else {
+			// Target is also IPv6, use as-is
+			targetIP = targetTCP.IP.String()
+		}
+	}
+
+	return fmt.Sprintf("PROXY %s %s %s %d %d\r\n",
+		protocol,
+		proxyInfo.SrcIP,
+		targetIP,
+		proxyInfo.SrcPort,
+		targetTCP.Port)
+}
+
+// buildProxyProtocolHeader creates a PROXY protocol v1 header
+func buildProxyProtocolHeader(clientAddr, targetAddr net.Addr) string {
+	clientTCP, ok := clientAddr.(*net.TCPAddr)
+	if !ok {
+		// Fallback for unknown address types
+		return "PROXY UNKNOWN\r\n"
+	}
+
+	targetTCP, ok := targetAddr.(*net.TCPAddr)
+	if !ok {
+		// Fallback for unknown address types
+		return "PROXY UNKNOWN\r\n"
+	}
+
+	// Determine protocol family based on client IP and normalize target IP accordingly
+	var protocol string
+	var targetIP string
+
+	if clientTCP.IP.To4() != nil {
+		// Client is IPv4, use TCP4 protocol
+		protocol = "TCP4"
+		if targetTCP.IP.To4() != nil {
+			// Target is also IPv4, use as-is
+			targetIP = targetTCP.IP.String()
+		} else {
+			// Target is IPv6, but we need IPv4 for consistent protocol family
+			// Use the IPv4 loopback if target is IPv6 loopback, otherwise use 127.0.0.1
+			if targetTCP.IP.IsLoopback() {
+				targetIP = "127.0.0.1"
+			} else {
+				// For non-loopback IPv6 targets, we could try to extract embedded IPv4
+				// or fall back to a sensible IPv4 address based on the target
+				targetIP = "127.0.0.1" // Safe fallback
+			}
+		}
+	} else {
+		// Client is IPv6, use TCP6 protocol
+		protocol = "TCP6"
+		if targetTCP.IP.To4() != nil {
+			// Target is IPv4, convert to IPv6 representation
+			targetIP = "::ffff:" + targetTCP.IP.String()
+		} else {
+			// Target is also IPv6, use as-is
+			targetIP = targetTCP.IP.String()
+		}
+	}
+
+	return fmt.Sprintf("PROXY %s %s %s %d %d\r\n",
+		protocol,
+		clientTCP.IP.String(),
+		targetIP,
+		clientTCP.Port,
+		targetTCP.Port)
+}
+
 // NewSNIProxy creates a new SNI proxy instance
 func NewSNIProxy(port int, remoteConfigURL, publicKey, localProxyAddr string, localProxyPort int, localOverrides []string, proxyProtocol bool, trustedUpstreams []string) (*SNIProxy, error) {
 	ctx, cancel := context.WithCancel(context.Background())
@@ -121,6 +435,20 @@ func NewSNIProxy(port int, remoteConfigURL, publicKey, localProxyAddr string, lo
 		localOverrides:   overridesMap,
 		activeTunnels:    make(map[string]*activeTunnel),
 		trustedUpstreams: trustedMap,
+		httpClient: &http.Client{
+			Timeout: 5 * time.Second,
+			Transport: &http.Transport{
+				MaxIdleConns:        100,
+				MaxIdleConnsPerHost: 10,
+				IdleConnTimeout:     90 * time.Second,
+			},
+		},
+		bufferPool: &sync.Pool{
+			New: func() interface{} {
+				buf := make([]byte, 32*1024)
+				return &buf
+			},
+		},
 	}
 
 	return proxy, nil
@@ -184,8 +512,46 @@ func (p *SNIProxy) acceptConnections() {
 			}
 		}
 
+		if p.activeConnections.Load() >= maxSNIConnections {
+			logger.Debug("Max concurrent SNI connections (%d) reached, rejecting connection from %s", maxSNIConnections, conn.RemoteAddr())
+			metrics.RecordSNIConnection("rejected_max_connections")
+			conn.Close()
+			continue
+		}
+
+		remoteHost, _, err := net.SplitHostPort(conn.RemoteAddr().String())
+		if err != nil {
+			remoteHost = conn.RemoteAddr().String()
+		}
+
+		counterVal, _ := p.perIPConnections.LoadOrStore(remoteHost, new(atomic.Int64))
+		perIPCounter := counterVal.(*atomic.Int64)
+		if perIPCounter.Load() >= maxSNIConnectionsPerIP {
+			logger.Debug("Max concurrent SNI connections per IP (%d) reached for %s, rejecting connection", maxSNIConnectionsPerIP, remoteHost)
+			metrics.RecordSNIConnection("rejected_per_ip_limit")
+			conn.Close()
+			continue
+		}
+		perIPCounter.Add(1)
+
+		p.activeConnections.Add(1)
+		metrics.RecordSNIActiveConnection(1)
 		p.wg.Add(1)
-		go p.handleConnection(conn)
+		go func() {
+			defer func() {
+				if perIPCounter.Add(-1) == 0 {
+					// Best-effort cleanup: only remove the map entry if it
+					// still holds this exact counter (a concurrent new
+					// connection from the same IP may have already bumped
+					// it back up via LoadOrStore, or replaced it after a
+					// prior race). Undercounting in that narrow race window
+					// just means one connection isn't rate-limited briefly,
+					// never unbounded growth.
+					p.perIPConnections.CompareAndDelete(remoteHost, counterVal)
+				}
+			}()
+			p.handleConnection(conn)
+		}()
 	}
 }
 
@@ -233,21 +599,29 @@ func (p *SNIProxy) extractSNI(conn net.Conn) (string, io.Reader, error) {
 func (p *SNIProxy) handleConnection(clientConn net.Conn) {
 	defer p.wg.Done()
 	defer clientConn.Close()
+	defer func() {
+		p.activeConnections.Add(-1)
+		metrics.RecordSNIActiveConnection(-1)
+	}()
+
+	metrics.RecordSNIConnection("accepted")
 
 	logger.Debug("Accepted connection from %s", clientConn.RemoteAddr())
 
 	// Check for PROXY protocol from trusted upstream
-	var proxyInfo *proxyproto.Info
+	var proxyInfo *ProxyProtocolInfo
 	var actualClientConn net.Conn = clientConn
 
 	if len(p.trustedUpstreams) > 0 {
 		var err error
-		proxyInfo, actualClientConn, err = proxyproto.ParseV1Header(clientConn, p.trustedUpstreams)
+		proxyInfo, actualClientConn, err = p.parseProxyProtocolHeader(clientConn)
 		if err != nil {
+			metrics.RecordSNIProxyProtocolParseError()
 			logger.Debug("Failed to parse PROXY protocol: %v", err)
 			return
 		}
 		if proxyInfo != nil {
+			metrics.RecordSNITrustedProxyEvent("proxy_protocol_parsed")
 			logger.Debug("Received PROXY protocol from trusted upstream: %s:%d -> %s:%d",
 				proxyInfo.SrcIP, proxyInfo.SrcPort, proxyInfo.DestIP, proxyInfo.DestPort)
 		} else {
@@ -264,11 +638,13 @@ func (p *SNIProxy) handleConnection(clientConn net.Conn) {
 	}
 
 	// Extract SNI hostname
+	clientHelloStart := time.Now()
 	hostname, clientReader, err := p.extractSNI(actualClientConn)
 	if err != nil {
 		logger.Debug("SNI extraction failed: %v", err)
 		return
 	}
+	metrics.RecordProxyTLSHandshake(time.Since(clientHelloStart).Seconds())
 
 	if hostname == "" {
 		log.Println("No SNI hostname found")
@@ -316,16 +692,18 @@ func (p *SNIProxy) handleConnection(clientConn net.Conn) {
 	defer targetConn.Close()
 
 	logger.Debug("Connected to target: %s:%d", route.TargetHost, route.TargetPort)
+	metrics.RecordActiveProxyConnection(1)
+	defer metrics.RecordActiveProxyConnection(-1)
 
 	// Send PROXY protocol header if enabled
 	if p.proxyProtocol {
 		var proxyHeader string
 		if proxyInfo != nil {
 			// Use original client info from PROXY protocol
-			proxyHeader = proxyproto.BuildV1HeaderFromInfo(proxyInfo, targetConn.LocalAddr())
+			proxyHeader = p.buildProxyProtocolHeaderFromInfo(proxyInfo, targetConn.LocalAddr())
 		} else {
 			// Use direct client connection info
-			proxyHeader = proxyproto.BuildV1Header(clientConn.RemoteAddr(), targetConn.LocalAddr())
+			proxyHeader = buildProxyProtocolHeader(clientConn.RemoteAddr(), targetConn.LocalAddr())
 		}
 		logger.Debug("Sending PROXY protocol header: %s", strings.TrimSpace(proxyHeader))
 
@@ -365,7 +743,7 @@ func (p *SNIProxy) handleConnection(clientConn net.Conn) {
 	}()
 
 	// Start bidirectional data transfer
-	p.pipe(actualClientConn, targetConn, clientReader)
+	p.pipe(hostname, actualClientConn, targetConn, clientReader)
 }
 
 // getRoute retrieves routing information for a hostname
@@ -373,6 +751,7 @@ func (p *SNIProxy) getRoute(hostname, clientAddr string) (*RouteRecord, error) {
 	// Check local overrides first
 	if _, isOverride := p.localOverrides[hostname]; isOverride {
 		logger.Debug("Local override matched for hostname: %s", hostname)
+		metrics.RecordProxyRouteLookup("local_override")
 		return &RouteRecord{
 			Hostname:   hostname,
 			TargetHost: p.localProxyAddr,
@@ -385,6 +764,7 @@ func (p *SNIProxy) getRoute(hostname, clientAddr string) (*RouteRecord, error) {
 	_, isLocal := p.localSNIs[hostname]
 	p.localSNIsLock.RUnlock()
 	if isLocal {
+		metrics.RecordProxyRouteLookup("local")
 		return &RouteRecord{
 			Hostname:   hostname,
 			TargetHost: p.localProxyAddr,
@@ -395,13 +775,16 @@ func (p *SNIProxy) getRoute(hostname, clientAddr string) (*RouteRecord, error) {
 	// Check cache first
 	if cached, found := p.cache.Get(hostname); found {
 		if cached == nil {
+			metrics.RecordProxyRouteLookup("cached_not_found")
 			return nil, nil // Cached negative result
 		}
 		logger.Debug("Cache hit for hostname: %s", hostname)
+		metrics.RecordProxyRouteLookup("cache_hit")
 		return cached.(*RouteRecord), nil
 	}
 
 	logger.Debug("Cache miss for hostname: %s, querying API", hostname)
+	metrics.RecordProxyRouteLookup("cache_miss")
 
 	// Query API with timeout
 	ctx, cancel := context.WithTimeout(p.ctx, 5*time.Second)
@@ -429,22 +812,28 @@ func (p *SNIProxy) getRoute(hostname, clientAddr string) (*RouteRecord, error) {
 	req.Header.Set("Content-Type", "application/json")
 
 	// Make HTTP request
-	client := &http.Client{Timeout: 5 * time.Second}
-	resp, err := client.Do(req)
+	apiStart := time.Now()
+	// Make HTTP request using reusable client
+	resp, err := p.httpClient.Do(req)
 	if err != nil {
+		metrics.RecordSNIRouteAPIRequest("error")
 		return nil, fmt.Errorf("API request failed: %w", err)
 	}
 	defer resp.Body.Close()
+	metrics.RecordSNIRouteAPILatency(time.Since(apiStart).Seconds())
 
 	if resp.StatusCode == http.StatusNotFound {
+		metrics.RecordSNIRouteAPIRequest("not_found")
 		// Cache negative result for shorter time (1 minute)
 		p.cache.Set(hostname, nil, 1*time.Minute)
 		return nil, nil
 	}
 
 	if resp.StatusCode != http.StatusOK {
+		metrics.RecordSNIRouteAPIRequest("error")
 		return nil, fmt.Errorf("API returned status %d", resp.StatusCode)
 	}
+	metrics.RecordSNIRouteAPIRequest("success")
 
 	// Parse response
 	var apiResponse RouteAPIResponse
@@ -500,8 +889,25 @@ func (p *SNIProxy) selectStickyEndpoint(clientAddr string, endpoints []string) s
 	return endpoints[index]
 }
 
+// bufferedReader hides any io.WriterTo the wrapped reader implements (e.g.
+// io.MultiReader, used by peekClientHello to replay the buffered
+// ClientHello ahead of the raw connection). Without this, io.CopyBuffer
+// bypasses the caller-supplied buffer entirely and lets WriteTo drive its
+// own, uncapped allocations - defeating the point of bufferPool.
+type bufferedReader struct {
+	io.Reader
+}
+
+// bufferedWriter hides any io.ReaderFrom the wrapped writer implements (e.g.
+// *net.TCPConn's splice/sendfile fast path). Without this, io.CopyBuffer
+// bypasses the caller-supplied buffer here too, so each copy allocates and
+// manages its own buffer regardless of what's pooled.
+type bufferedWriter struct {
+	io.Writer
+}
+
 // pipe handles bidirectional data transfer between connections
-func (p *SNIProxy) pipe(clientConn, targetConn net.Conn, clientReader io.Reader) {
+func (p *SNIProxy) pipe(hostname string, clientConn, targetConn net.Conn, clientReader io.Reader) {
 	var wg sync.WaitGroup
 	wg.Add(2)
 
@@ -520,9 +926,16 @@ func (p *SNIProxy) pipe(clientConn, targetConn net.Conn, clientReader io.Reader)
 		defer wg.Done()
 		defer closeConns()
 
-		// Use a large buffer for better performance
-		buf := make([]byte, 32*1024)
-		_, err := io.CopyBuffer(targetConn, clientReader, buf)
+		// Get buffer from pool and return when done
+		bufPtr := p.bufferPool.Get().(*[]byte)
+		defer func() {
+			// Clear buffer before returning to pool to prevent data leakage
+			clear(*bufPtr)
+			p.bufferPool.Put(bufPtr)
+		}()
+
+		bytesCopied, err := io.CopyBuffer(bufferedWriter{targetConn}, bufferedReader{clientReader}, *bufPtr)
+		metrics.RecordProxyBytesTransmitted("client_to_target", bytesCopied)
 		if err != nil && err != io.EOF {
 			logger.Debug("Copy client->target error: %v", err)
 		}
@@ -533,9 +946,16 @@ func (p *SNIProxy) pipe(clientConn, targetConn net.Conn, clientReader io.Reader)
 		defer wg.Done()
 		defer closeConns()
 
-		// Use a large buffer for better performance
-		buf := make([]byte, 32*1024)
-		_, err := io.CopyBuffer(clientConn, targetConn, buf)
+		// Get buffer from pool and return when done
+		bufPtr := p.bufferPool.Get().(*[]byte)
+		defer func() {
+			// Clear buffer before returning to pool to prevent data leakage
+			clear(*bufPtr)
+			p.bufferPool.Put(bufPtr)
+		}()
+
+		bytesCopied, err := io.CopyBuffer(bufferedWriter{clientConn}, bufferedReader{targetConn}, *bufPtr)
+		metrics.RecordProxyBytesTransmitted("target_to_client", bytesCopied)
 		if err != nil && err != io.EOF {
 			logger.Debug("Copy target->client error: %v", err)
 		}

proxy/proxy_test.go

@@ -3,8 +3,6 @@ package proxy
 import (
 	"net"
 	"testing"
-
-	"github.com/fosrl/gerbil/proxyproto"
 )
 
 func TestBuildProxyProtocolHeader(t *testing.T) {
@@ -58,7 +56,7 @@ func TestBuildProxyProtocolHeader(t *testing.T) {
 				t.Fatalf("Failed to resolve target address: %v", err)
 			}
 
-			result := proxyproto.BuildV1Header(clientTCP, targetTCP)
+			result := buildProxyProtocolHeader(clientTCP, targetTCP)
 			if result != tt.expected {
 				t.Errorf("Expected %q, got %q", tt.expected, result)
 			}
@@ -71,7 +69,7 @@ func TestBuildProxyProtocolHeaderUnknownType(t *testing.T) {
 	clientAddr := &net.UDPAddr{IP: net.ParseIP("192.168.1.100"), Port: 12345}
 	targetAddr := &net.UDPAddr{IP: net.ParseIP("10.0.0.1"), Port: 443}
 
-	result := proxyproto.BuildV1Header(clientAddr, targetAddr)
+	result := buildProxyProtocolHeader(clientAddr, targetAddr)
 	expected := "PROXY UNKNOWN\r\n"
 
 	if result != expected {
@@ -80,8 +78,13 @@ func TestBuildProxyProtocolHeaderUnknownType(t *testing.T) {
 }
 
 func TestBuildProxyProtocolHeaderFromInfo(t *testing.T) {
+	proxy, err := NewSNIProxy(8443, "", "", "127.0.0.1", 443, nil, true, nil)
+	if err != nil {
+		t.Fatalf("Failed to create SNI proxy: %v", err)
+	}
+
 	// Test IPv4 case
-	info := &proxyproto.Info{
+	proxyInfo := &ProxyProtocolInfo{
 		Protocol: "TCP4",
 		SrcIP:    "10.0.0.1",
 		DestIP:   "192.168.1.100",
@@ -90,7 +93,7 @@ func TestBuildProxyProtocolHeaderFromInfo(t *testing.T) {
 	}
 
 	targetAddr, _ := net.ResolveTCPAddr("tcp", "127.0.0.1:8080")
-	header := proxyproto.BuildV1HeaderFromInfo(info, targetAddr)
+	header := proxy.buildProxyProtocolHeaderFromInfo(proxyInfo, targetAddr)
 
 	expected := "PROXY TCP4 10.0.0.1 127.0.0.1 12345 8080\r\n"
 	if header != expected {
@@ -98,7 +101,7 @@ func TestBuildProxyProtocolHeaderFromInfo(t *testing.T) {
 	}
 
 	// Test IPv6 case
-	info = &proxyproto.Info{
+	proxyInfo = &ProxyProtocolInfo{
 		Protocol: "TCP6",
 		SrcIP:    "2001:db8::1",
 		DestIP:   "2001:db8::2",
@@ -107,99 +110,10 @@ func TestBuildProxyProtocolHeaderFromInfo(t *testing.T) {
 	}
 
 	targetAddr, _ = net.ResolveTCPAddr("tcp6", "[::1]:8080")
-	header = proxyproto.BuildV1HeaderFromInfo(info, targetAddr)
+	header = proxy.buildProxyProtocolHeaderFromInfo(proxyInfo, targetAddr)
 
 	expected = "PROXY TCP6 2001:db8::1 ::1 12345 8080\r\n"
 	if header != expected {
 		t.Errorf("Expected header '%s', got '%s'", expected, header)
 	}
 }
-
-func TestParseV2UDPHeader(t *testing.T) {
-	// Build a minimal PROXY v2 header for IPv4 UDP
-	// Magic (12) + ver/cmd (1) + fam/proto (1) + len (2) + src IP (4) + dst IP (4) + src port (2) + dst port (2) = 28 bytes
-	header := []byte{
-		// Magic signature
-		0x0D, 0x0A, 0x0D, 0x0A, 0x00, 0x0D, 0x0A, 0x51, 0x55, 0x49, 0x54, 0x0A,
-		// Version 2 (0x2x), PROXY command (0x01)
-		0x21,
-		// AF_INET (0x1x), DGRAM/UDP (0x02)
-		0x12,
-		// Address block length: 12 bytes (4+4+2+2)
-		0x00, 0x0C,
-		// Source IP: 192.168.1.100
-		192, 168, 1, 100,
-		// Destination IP: 10.0.0.1
-		10, 0, 0, 1,
-		// Source port: 4500
-		0x11, 0x94,
-		// Destination port: 21820
-		0x55, 0x3C,
-	}
-
-	// Append a fake application payload
-	payload := []byte{0x01, 0x02, 0x03}
-	data := append(header, payload...)
-
-	info, remaining, ok := proxyproto.ParseV2UDPHeader(data)
-	if !ok {
-		t.Fatal("Expected ParseV2UDPHeader to return ok=true")
-	}
-	if info == nil {
-		t.Fatal("Expected non-nil Info")
-	}
-	if info.Protocol != "UDP4" {
-		t.Errorf("Expected protocol UDP4, got %s", info.Protocol)
-	}
-	if info.SrcIP != "192.168.1.100" {
-		t.Errorf("Expected SrcIP 192.168.1.100, got %s", info.SrcIP)
-	}
-	if info.DestIP != "10.0.0.1" {
-		t.Errorf("Expected DestIP 10.0.0.1, got %s", info.DestIP)
-	}
-	if info.SrcPort != 4500 {
-		t.Errorf("Expected SrcPort 4500, got %d", info.SrcPort)
-	}
-	if info.DestPort != 21820 {
-		t.Errorf("Expected DestPort 21820, got %d", info.DestPort)
-	}
-	if len(remaining) != len(payload) {
-		t.Errorf("Expected %d remaining bytes, got %d", len(payload), len(remaining))
-	}
-}
-
-func TestParseV2UDPHeaderNoHeader(t *testing.T) {
-	// Data that does NOT start with v2 magic should be returned as-is
-	data := []byte{0x01, 0x02, 0x03}
-	info, remaining, ok := proxyproto.ParseV2UDPHeader(data)
-	if ok {
-		t.Error("Expected ok=false for non-v2 data")
-	}
-	if info != nil {
-		t.Error("Expected nil Info for non-v2 data")
-	}
-	if len(remaining) != len(data) {
-		t.Errorf("Expected remaining to equal original data length %d, got %d", len(data), len(remaining))
-	}
-}
-
-func TestIsV2Header(t *testing.T) {
-	valid := []byte{
-		0x0D, 0x0A, 0x0D, 0x0A, 0x00, 0x0D, 0x0A, 0x51, 0x55, 0x49, 0x54, 0x0A,
-		// extra bytes beyond the magic
-		0x21, 0x12,
-	}
-	if !proxyproto.IsV2Header(valid) {
-		t.Error("Expected IsV2Header=true for valid magic")
-	}
-
-	invalid := []byte{0x01, 0x02, 0x03}
-	if proxyproto.IsV2Header(invalid) {
-		t.Error("Expected IsV2Header=false for non-magic data")
-	}
-
-	tooShort := []byte{0x0D, 0x0A}
-	if proxyproto.IsV2Header(tooShort) {
-		t.Error("Expected IsV2Header=false for too-short data")
-	}
-}

proxyproto/proxyproto.go

@@ -1,370 +0,0 @@
-// Package proxyproto provides shared PROXY protocol v1 (TCP) and v2 (UDP) parsing
-// and header building utilities used by both the SNI proxy and UDP relay components.
-package proxyproto
-
-import (
-	"bytes"
-	"encoding/binary"
-	"fmt"
-	"io"
-	"net"
-	"strconv"
-	"strings"
-	"time"
-
-	"github.com/fosrl/gerbil/logger"
-)
-
-// v2Signature is the 12-byte magic prefix for PROXY protocol v2 headers.
-var v2Signature = []byte{
-	0x0D, 0x0A, 0x0D, 0x0A, 0x00, 0x0D, 0x0A, 0x51, 0x55, 0x49, 0x54, 0x0A,
-}
-
-// Info holds information parsed from an incoming PROXY protocol header (v1 or v2).
-type Info struct {
-	Protocol string // e.g. "TCP4", "TCP6", "UDP4", "UDP6"
-	SrcIP    string
-	DestIP   string
-	SrcPort  int
-	DestPort int
-}
-
-// Conn wraps a net.Conn so that reads are satisfied from a pre-pended buffered
-// reader first (remaining bytes after PROXY header parsing) and then from the
-// underlying connection. All other net.Conn methods are forwarded unchanged.
-type Conn struct {
-	net.Conn
-	Reader io.Reader
-}
-
-// Read satisfies net.Conn, draining the buffered reader before falling through
-// to the underlying connection.
-func (c *Conn) Read(b []byte) (int, error) {
-	return c.Reader.Read(b)
-}
-
-// IsV2Header returns true when data begins with the 12-byte PROXY protocol v2
-// magic signature.
-func IsV2Header(data []byte) bool {
-	if len(data) < 12 {
-		return false
-	}
-	return bytes.Equal(data[:12], v2Signature)
-}
-
-// ParseV2UDPHeader tries to parse a PROXY protocol v2 header from the front of
-// a UDP datagram payload.
-//
-// Three return values are provided:
-//   - *Info  – filled when a PROXY command header was parsed successfully; nil
-//     for a LOCAL command or unrecognised address family.
-//   - []byte – the remaining payload that follows the header (the actual
-//     application data).
-//   - bool   – true when a v2 header was detected (and consumed), false when
-//     no v2 magic is present and data should be treated as-is.
-func ParseV2UDPHeader(data []byte) (*Info, []byte, bool) {
-	if !IsV2Header(data) {
-		return nil, data, false
-	}
-
-	// Minimum fixed header size: 12 (magic) + 1 (ver/cmd) + 1 (fam/proto) + 2 (len) = 16
-	if len(data) < 16 {
-		return nil, data, false
-	}
-
-	// Byte 12: version (high nibble) + command (low nibble)
-	versionCmd := data[12]
-	version := (versionCmd >> 4) & 0x0F
-	command := versionCmd & 0x0F
-
-	if version != 2 {
-		return nil, data, false
-	}
-
-	// Byte 13: address family (high nibble) + transport protocol (low nibble)
-	familyProto := data[13]
-	family := (familyProto >> 4) & 0x0F
-	protocol := familyProto & 0x0F
-
-	// Bytes 14-15: length of the address block that follows, big-endian
-	addrLen := int(binary.BigEndian.Uint16(data[14:16]))
-	totalHeaderLen := 16 + addrLen
-
-	if len(data) < totalHeaderLen {
-		// Truncated packet – signal that a header was detected but is malformed
-		return nil, data, false
-	}
-
-	payload := data[totalHeaderLen:]
-
-	// LOCAL command (0) carries no address information.
-	if command == 0 {
-		return nil, payload, true
-	}
-
-	if command != 1 {
-		// Unknown command – consume the header and return no info
-		return nil, payload, true
-	}
-
-	addrBlock := data[16:totalHeaderLen]
-
-	var (
-		srcIP, destIP net.IP
-		srcPort       uint16
-		destPort      uint16
-		protocolStr   string
-	)
-
-	switch {
-	case family == 1 && protocol == 1: // AF_INET / STREAM  (TCP over IPv4)
-		if len(addrBlock) < 12 {
-			return nil, payload, false
-		}
-		srcIP = net.IP(addrBlock[0:4])
-		destIP = net.IP(addrBlock[4:8])
-		srcPort = binary.BigEndian.Uint16(addrBlock[8:10])
-		destPort = binary.BigEndian.Uint16(addrBlock[10:12])
-		protocolStr = "TCP4"
-
-	case family == 1 && protocol == 2: // AF_INET / DGRAM   (UDP over IPv4)
-		if len(addrBlock) < 12 {
-			return nil, payload, false
-		}
-		srcIP = net.IP(addrBlock[0:4])
-		destIP = net.IP(addrBlock[4:8])
-		srcPort = binary.BigEndian.Uint16(addrBlock[8:10])
-		destPort = binary.BigEndian.Uint16(addrBlock[10:12])
-		protocolStr = "UDP4"
-
-	case family == 2 && protocol == 1: // AF_INET6 / STREAM (TCP over IPv6)
-		if len(addrBlock) < 36 {
-			return nil, payload, false
-		}
-		srcIP = net.IP(addrBlock[0:16])
-		destIP = net.IP(addrBlock[16:32])
-		srcPort = binary.BigEndian.Uint16(addrBlock[32:34])
-		destPort = binary.BigEndian.Uint16(addrBlock[34:36])
-		protocolStr = "TCP6"
-
-	case family == 2 && protocol == 2: // AF_INET6 / DGRAM  (UDP over IPv6)
-		if len(addrBlock) < 36 {
-			return nil, payload, false
-		}
-		srcIP = net.IP(addrBlock[0:16])
-		destIP = net.IP(addrBlock[16:32])
-		srcPort = binary.BigEndian.Uint16(addrBlock[32:34])
-		destPort = binary.BigEndian.Uint16(addrBlock[34:36])
-		protocolStr = "UDP6"
-
-	default:
-		// UNSPEC or AF_UNIX – consume the header, no address info available
-		return nil, payload, true
-	}
-
-	info := &Info{
-		Protocol: protocolStr,
-		SrcIP:    srcIP.String(),
-		DestIP:   destIP.String(),
-		SrcPort:  int(srcPort),
-		DestPort: int(destPort),
-	}
-	return info, payload, true
-}
-
-// ParseV1Header attempts to parse a PROXY protocol v1 (text) header from the
-// given TCP connection.
-//
-// The function first checks whether the remote address appears in
-// trustedUpstreams. If it does not, it returns (nil, conn, nil) and the caller
-// should treat the connection as a plain (non-proxied) connection.
-//
-// When a trusted upstream is detected the function reads up to 512 bytes,
-// locates the CRLF-terminated header line, and parses the proxy information.
-// Whatever bytes were consumed (including any data beyond the header line) are
-// re-prepended via a *Conn wrapper so that subsequent reads by the caller are
-// transparent.
-//
-// Return values:
-//   - *Info    – non-nil when a valid PROXY header was parsed.
-//   - net.Conn – always a valid connection (possibly a *Conn wrapper).
-//   - error    – non-nil only on hard failures (e.g. bad port numbers).
-func ParseV1Header(conn net.Conn, trustedUpstreams map[string]struct{}) (*Info, net.Conn, error) {
-	remoteHost, _, err := net.SplitHostPort(conn.RemoteAddr().String())
-	if err != nil {
-		return nil, conn, fmt.Errorf("failed to parse remote address: %w", err)
-	}
-
-	if _, isTrusted := trustedUpstreams[remoteHost]; !isTrusted {
-		return nil, conn, nil
-	}
-
-	// Give the upstream 5 s to deliver the PROXY header before timing out.
-	if err := conn.SetReadDeadline(time.Now().Add(5 * time.Second)); err != nil {
-		return nil, conn, fmt.Errorf("failed to set read deadline: %w", err)
-	}
-
-	// The PROXY v1 spec mandates the header fits in 108 bytes; 512 is generous.
-	buffer := make([]byte, 512)
-	n, err := conn.Read(buffer)
-	if err != nil {
-		logger.Debug("Could not read from trusted upstream %s, treating as regular connection: %v", remoteHost, err)
-		if clearErr := conn.SetReadDeadline(time.Time{}); clearErr != nil {
-			logger.Debug("Failed to clear read deadline: %v", clearErr)
-		}
-		return nil, conn, nil
-	}
-
-	// Locate the CRLF that terminates the PROXY header line.
-	headerEnd := bytes.Index(buffer[:n], []byte("\r\n"))
-	if headerEnd == -1 {
-		logger.Debug("No PROXY protocol header from trusted upstream %s, treating as regular TLS connection", remoteHost)
-		if clearErr := conn.SetReadDeadline(time.Time{}); clearErr != nil {
-			logger.Debug("Failed to clear read deadline: %v", clearErr)
-		}
-		newReader := io.MultiReader(bytes.NewReader(buffer[:n]), conn)
-		return nil, &Conn{Conn: conn, Reader: newReader}, nil
-	}
-
-	headerLine := string(buffer[:headerEnd])
-	remainingData := buffer[headerEnd+2 : n]
-
-	parts := strings.Fields(headerLine)
-
-	// Handle "PROXY UNKNOWN" – upstream knows the real source but we don't need it.
-	if len(parts) == 2 && parts[0] == "PROXY" && parts[1] == "UNKNOWN" {
-		if clearErr := conn.SetReadDeadline(time.Time{}); clearErr != nil {
-			logger.Debug("Failed to clear read deadline: %v", clearErr)
-		}
-		var newConn net.Conn
-		if len(remainingData) > 0 {
-			newConn = &Conn{Conn: conn, Reader: io.MultiReader(bytes.NewReader(remainingData), conn)}
-		} else {
-			newConn = conn
-		}
-		return nil, newConn, nil
-	}
-
-	if len(parts) != 6 || parts[0] != "PROXY" {
-		// Malformed line from a trusted upstream – re-prepend everything and
-		// let the caller deal with it as a plain TLS connection.
-		logger.Debug("Invalid PROXY protocol from trusted upstream %s, treating as regular TLS connection: %s", remoteHost, headerLine)
-		if clearErr := conn.SetReadDeadline(time.Time{}); clearErr != nil {
-			logger.Debug("Failed to clear read deadline: %v", clearErr)
-		}
-		newReader := io.MultiReader(bytes.NewReader(buffer[:n]), conn)
-		return nil, &Conn{Conn: conn, Reader: newReader}, nil
-	}
-
-	protocol := parts[1]
-	srcIP := parts[2]
-	destIP := parts[3]
-
-	srcPort, err := strconv.Atoi(parts[4])
-	if err != nil {
-		return nil, conn, fmt.Errorf("invalid source port in PROXY header: %s", parts[4])
-	}
-	destPort, err := strconv.Atoi(parts[5])
-	if err != nil {
-		return nil, conn, fmt.Errorf("invalid destination port in PROXY header: %s", parts[5])
-	}
-
-	// Re-assemble a reader that returns any bytes read beyond the header first.
-	var newReader io.Reader
-	if len(remainingData) > 0 {
-		newReader = io.MultiReader(bytes.NewReader(remainingData), conn)
-	} else {
-		newReader = conn
-	}
-	wrappedConn := &Conn{Conn: conn, Reader: newReader}
-
-	if clearErr := conn.SetReadDeadline(time.Time{}); clearErr != nil {
-		return nil, conn, fmt.Errorf("failed to clear read deadline: %w", clearErr)
-	}
-
-	info := &Info{
-		Protocol: protocol,
-		SrcIP:    srcIP,
-		DestIP:   destIP,
-		SrcPort:  srcPort,
-		DestPort: destPort,
-	}
-	return info, wrappedConn, nil
-}
-
-// BuildV1Header constructs a PROXY protocol v1 header string from two TCP
-// addresses, normalising the protocol family so that v1's constraint of a
-// single family per header is satisfied.
-func BuildV1Header(clientAddr, targetAddr net.Addr) string {
-	clientTCP, ok := clientAddr.(*net.TCPAddr)
-	if !ok {
-		return "PROXY UNKNOWN\r\n"
-	}
-	targetTCP, ok := targetAddr.(*net.TCPAddr)
-	if !ok {
-		return "PROXY UNKNOWN\r\n"
-	}
-
-	var protocol, targetIP string
-
-	if clientTCP.IP.To4() != nil {
-		// IPv4 client
-		protocol = "TCP4"
-		if targetTCP.IP.To4() != nil {
-			targetIP = targetTCP.IP.String()
-		} else if targetTCP.IP.IsLoopback() {
-			targetIP = "127.0.0.1"
-		} else {
-			targetIP = "127.0.0.1" // safe fallback for mixed-family
-		}
-	} else {
-		// IPv6 client
-		protocol = "TCP6"
-		if targetTCP.IP.To4() != nil {
-			targetIP = "::ffff:" + targetTCP.IP.String()
-		} else {
-			targetIP = targetTCP.IP.String()
-		}
-	}
-
-	return fmt.Sprintf("PROXY %s %s %s %d %d\r\n",
-		protocol, clientTCP.IP.String(), targetIP, clientTCP.Port, targetTCP.Port)
-}
-
-// BuildV1HeaderFromInfo constructs a PROXY protocol v1 header string using a
-// previously-parsed *Info (i.e. when this server itself sits behind an
-// upstream proxy) and the target TCP address.
-func BuildV1HeaderFromInfo(info *Info, targetAddr net.Addr) string {
-	targetTCP, ok := targetAddr.(*net.TCPAddr)
-	if !ok {
-		return "PROXY UNKNOWN\r\n"
-	}
-
-	srcIP := net.ParseIP(info.SrcIP)
-	if srcIP == nil {
-		return "PROXY UNKNOWN\r\n"
-	}
-
-	var protocol, targetIP string
-
-	if srcIP.To4() != nil {
-		protocol = "TCP4"
-		if targetTCP.IP.To4() != nil {
-			targetIP = targetTCP.IP.String()
-		} else if targetTCP.IP.IsLoopback() {
-			targetIP = "127.0.0.1"
-		} else {
-			targetIP = "127.0.0.1"
-		}
-	} else {
-		protocol = "TCP6"
-		if targetTCP.IP.To4() != nil {
-			targetIP = "::ffff:" + targetTCP.IP.String()
-		} else {
-			targetIP = targetTCP.IP.String()
-		}
-	}
-
-	return fmt.Sprintf("PROXY %s %s %s %d %d\r\n",
-		protocol, info.SrcIP, targetIP, info.SrcPort, targetTCP.Port)
-}

relay/relay.go

@@ -1,6 +1,7 @@
 package relay
 
 import (
+	"bufio"
 	"bytes"
 	"context"
 	"encoding/binary"
@@ -9,18 +10,22 @@ import (
 	"io"
 	"net"
 	"net/http"
+	"os"
 	"runtime"
-	"strings"
+	"strconv"
 	"sync"
+	"sync/atomic"
 	"time"
 
+	"github.com/fosrl/gerbil/internal/metrics"
 	"github.com/fosrl/gerbil/logger"
-	"github.com/fosrl/gerbil/proxyproto"
 	"golang.org/x/crypto/chacha20poly1305"
 	"golang.org/x/crypto/curve25519"
 	"golang.zx2c4.com/wireguard/wgctrl/wgtypes"
 )
 
+const relayIfname = "relay"
+
 type EncryptedHolePunchMessage struct {
 	EphemeralPublicKey string `json:"ephemeralPublicKey"`
 	Nonce              []byte `json:"nonce"`
@@ -58,8 +63,31 @@ type PeerDestination struct {
 }
 
 type DestinationConn struct {
-	conn     *net.UDPConn
-	lastUsed time.Time
+	conn *net.UDPConn
+	// lastUsed is unix nanoseconds, read/written via atomic ops since it's
+	// touched from packet workers and the response goroutine concurrently
+	// with no lock, and is also scanned for LRU eviction below.
+	lastUsed atomic.Int64
+}
+
+// defaultMaxUDPConnections caps the number of concurrent per-peer outbound
+// UDP sockets the relay will keep open in s.connections. Without a cap, a
+// burst of peer churn creates sockets faster than the 5-minute idle cleanup
+// can reap them, exhausting the host's ephemeral port range or fd ulimit.
+// That surfaces as "dial udp ...: resource temporarily unavailable" on every
+// subsequent packet and pegs the CPU logging the flood (outage 2026-07-03,
+// recurrence 2026-07-05). Overridable via GERBIL_MAX_UDP_CONNECTIONS.
+const defaultMaxUDPConnections = 8192
+
+var maxUDPConnections = loadMaxUDPConnections()
+
+func loadMaxUDPConnections() int64 {
+	if v := os.Getenv("GERBIL_MAX_UDP_CONNECTIONS"); v != "" {
+		if n, err := strconv.ParseInt(v, 10, 64); err == nil && n > 0 {
+			return n
+		}
+	}
+	return defaultMaxUDPConnections
 }
 
 // Type for storing WireGuard handshake information
@@ -136,6 +164,24 @@ const (
 	WireGuardMessageTypeTransportData       = 4
 )
 
+// cachedEndpointState holds the last-known endpoint fields used for change detection.
+// Timestamp is intentionally excluded since it always changes.
+type cachedEndpointState struct {
+	OlmID     string
+	NewtID    string
+	Token     string
+	IP        string
+	Port      int
+	PublicKey string
+}
+
+// cachedEndpointEntry wraps cachedEndpointState with a timestamp so the cache
+// can be expired after a short TTL even when the endpoint fields are unchanged.
+type cachedEndpointEntry struct {
+	state    cachedEndpointState
+	cachedAt time.Time
+}
+
 // --- End Types ---
 
 // bufferPool allows reusing buffers to reduce allocations.
@@ -152,14 +198,20 @@ type UDPProxyServer struct {
 	conn          *net.UDPConn
 	proxyMappings sync.Map // map[string]ProxyMapping where key is "ip:port"
 	connections   sync.Map // map[string]*DestinationConn where key is destination "ip:port"
-	privateKey    wgtypes.Key
-	packetChan    chan Packet
-	ctx           context.Context
-	cancel        context.CancelFunc
+	// connectionCount mirrors len(connections) without an O(n) sync.Map walk,
+	// so the cap check in getOrCreateConnection is cheap on the hot path.
+	connectionCount atomic.Int64
+	privateKey      wgtypes.Key
+	packetChan      chan Packet
+	ctx             context.Context
+	cancel          context.CancelFunc
 
 	// Session tracking for WireGuard peers
 	// Key format: "senderIndex:receiverIndex"
 	wgSessions sync.Map
+	// Session index for O(1) lookup by receiver index
+	// Key: receiverIndex (uint32), Value: *WireGuardSession
+	sessionsByReceiverIndex sync.Map
 	// Communication pattern tracking for rebuilding sessions
 	// Key format: "clientIP:clientPort-destIP:destPort"
 	commPatterns sync.Map
@@ -168,61 +220,42 @@ type UDPProxyServer struct {
 	// Cache for resolved UDP addresses to avoid per-packet DNS lookups
 	// Key: "ip:port" string, Value: *net.UDPAddr
 	addrCache sync.Map
+	// lastEndpointCache stores the last-known endpoint state per client (key: olmId:newtId)
+	// used to skip redundant HTTP notifications when nothing has changed.
+	lastEndpointCache sync.Map
+	// notifyChan is the async queue for hole-punch endpoint notifications.
+	// Dedicated notifier workers drain this channel and perform the HTTP call.
+	notifyChan chan ClientEndpoint
 	// ReachableAt is the URL where this server can be reached
 	ReachableAt string
-
-	// proxyProtocol enables PROXY protocol v2 header parsing for incoming UDP packets.
-	// When enabled, packets from trustedUpstreams that carry a v2 header will have
-	// their source address overridden with the address reported in the header.
-	proxyProtocol    bool
-	trustedUpstreams map[string]struct{}
 }
 
 // NewUDPProxyServer initializes the server with a buffered packet channel and derived context.
-//
-// proxyProtocol enables PROXY protocol v2 parsing for datagrams arriving from
-// any address listed in trustedUpstreams (plain IPs or resolvable hostnames).
-// When a trusted datagram carries a v2 header its source address is replaced
-// with the address carried inside the header before further processing, so that
-// hole-punch endpoints reflect the original client IP rather than the load
-// balancer's address.
-func NewUDPProxyServer(parentCtx context.Context, addr, serverURL string, privateKey wgtypes.Key, reachableAt string, proxyProtocol bool, trustedUpstreams []string) *UDPProxyServer {
+func NewUDPProxyServer(parentCtx context.Context, addr, serverURL string, privateKey wgtypes.Key, reachableAt string) *UDPProxyServer {
 	ctx, cancel := context.WithCancel(parentCtx)
-
-	trustedMap := make(map[string]struct{})
-	for _, upstream := range trustedUpstreams {
-		upstream = strings.TrimSpace(upstream)
-		if upstream == "" {
-			continue
-		}
-		trustedMap[upstream] = struct{}{}
-		// Also resolve any hostnames to their current IPs so we can match by IP.
-		if ips, err := net.LookupIP(upstream); err == nil {
-			for _, ip := range ips {
-				trustedMap[ip.String()] = struct{}{}
-			}
-		}
-	}
-
 	return &UDPProxyServer{
-		addr:             addr,
-		serverURL:        serverURL,
-		privateKey:       privateKey,
-		packetChan:       make(chan Packet, 50000), // Increased from 1000 to handle high throughput
-		ReachableAt:      reachableAt,
-		ctx:              ctx,
-		cancel:           cancel,
-		proxyProtocol:    proxyProtocol,
-		trustedUpstreams: trustedMap,
+		addr:        addr,
+		serverURL:   serverURL,
+		privateKey:  privateKey,
+		packetChan:  make(chan Packet, 50000), // Increased from 1000 to handle high throughput
+		notifyChan:  make(chan ClientEndpoint, 1000),
+		ReachableAt: reachableAt,
+		ctx:         ctx,
+		cancel:      cancel,
 	}
 }
 
 // Start sets up the UDP listener, worker pool, and begins reading packets.
 func (s *UDPProxyServer) Start() error {
-	// Fetch initial mappings.
-	if err := s.fetchInitialMappings(); err != nil {
-		return fmt.Errorf("failed to fetch initial mappings: %v", err)
-	}
+	// Fetch initial mappings asynchronously so a large (potentially 100MB+)
+	// response does not block the UDP listener from coming up. Any packets
+	// arriving for unknown mappings before the load completes will simply
+	// log and be repopulated via the hole-punch path.
+	go func() {
+		if err := s.fetchInitialMappings(); err != nil {
+			logger.Error("Failed to fetch initial mappings: %v", err)
+		}
+	}()
 
 	udpAddr, err := net.ResolveUDPAddr("udp", s.addr)
 	if err != nil {
@@ -264,6 +297,11 @@ func (s *UDPProxyServer) Start() error {
 	// Start the hole punch rate limiter cleanup routine
 	go s.cleanupHolePunchRateLimiter()
 
+	// Start async endpoint notifier workers (HTTP calls off the hot path)
+	for i := 0; i < 5; i++ {
+		go s.endpointNotifierWorker()
+	}
+
 	return nil
 }
 
@@ -321,48 +359,17 @@ func (s *UDPProxyServer) readPackets() {
 // packetWorker processes incoming packets from the channel.
 func (s *UDPProxyServer) packetWorker() {
 	for packet := range s.packetChan {
-		// effectiveData and effectiveAddr represent the application-layer payload
-		// and the true originating address.  They start as the raw UDP values and
-		// may be updated below when a PROXY protocol v2 header is present.
-		effectiveData := packet.data[:packet.n]
-		effectiveAddr := packet.remoteAddr
-
-		// ---------- PROXY protocol v2 (UDP) ------------------------------------
-		// If proxy protocol is enabled and this datagram arrives from a trusted
-		// upstream (e.g. a load balancer), attempt to parse the v2 header so
-		// that we use the original client address for hole-punch registration and
-		// WireGuard session tracking rather than the load balancer's address.
-		if s.proxyProtocol && len(s.trustedUpstreams) > 0 {
-			remoteHost := packet.remoteAddr.IP.String()
-			if _, trusted := s.trustedUpstreams[remoteHost]; trusted {
-				if info, payload, ok := proxyproto.ParseV2UDPHeader(effectiveData); ok {
-					if info != nil {
-						// Override source address with what the proxy reported.
-						if srcIP := net.ParseIP(info.SrcIP); srcIP != nil {
-							effectiveAddr = &net.UDPAddr{
-								IP:   srcIP,
-								Port: info.SrcPort,
-							}
-							logger.Debug("PROXY protocol v2: overriding source %s → %s:%d",
-								packet.remoteAddr, info.SrcIP, info.SrcPort)
-						}
-					}
-					// Always advance past the header so the remainder is treated
-					// as the real application payload.
-					effectiveData = payload
-				}
-			}
-		}
-		// -----------------------------------------------------------------------
-
-		// Determine packet type by inspecting the first byte of the (possibly
-		// stripped) application payload.
-		if len(effectiveData) > 0 && effectiveData[0] >= 1 && effectiveData[0] <= 4 {
+		// Determine packet type by inspecting the first byte.
+		if packet.n > 0 && packet.data[0] >= 1 && packet.data[0] <= 4 {
+			metrics.RecordUDPPacket(relayIfname, "wireguard", "in")
+			metrics.RecordUDPPacketSize(relayIfname, "wireguard", float64(packet.n))
 			// Process as a WireGuard packet.
-			s.handleWireGuardPacket(effectiveData, effectiveAddr)
+			s.handleWireGuardPacket(packet.data, packet.remoteAddr)
 		} else {
+			metrics.RecordUDPPacket(relayIfname, "hole_punch", "in")
+			metrics.RecordUDPPacketSize(relayIfname, "hole_punch", float64(packet.n))
 			// Rate limit: allow at most 2 hole punch messages per IP:Port per second
-			rateLimitKey := effectiveAddr.String()
+			rateLimitKey := packet.remoteAddr.String()
 			entryVal, _ := s.holePunchRateLimiter.LoadOrStore(rateLimitKey, &holePunchRateLimitEntry{
 				windowStart: time.Now(),
 			})
@@ -378,14 +385,16 @@ func (s *UDPProxyServer) packetWorker() {
 			rlEntry.mu.Unlock()
 			if !allowed {
 				// logger.Debug("Rate limiting hole punch message from %s", rateLimitKey)
+				metrics.RecordHolePunchEvent(relayIfname, "rate_limited")
 				bufferPool.Put(packet.data[:1500])
 				continue
 			}
 
 			// Process as an encrypted hole punch message
 			var encMsg EncryptedHolePunchMessage
-			if err := json.Unmarshal(effectiveData, &encMsg); err != nil {
+			if err := json.Unmarshal(packet.data, &encMsg); err != nil {
 				logger.Error("Error unmarshaling encrypted message: %v", err)
+				metrics.RecordHolePunchEvent(relayIfname, "error")
 				// Return the buffer to the pool for reuse and continue with next packet
 				bufferPool.Put(packet.data[:1500])
 				continue
@@ -393,6 +402,7 @@ func (s *UDPProxyServer) packetWorker() {
 
 			if encMsg.EphemeralPublicKey == "" {
 				logger.Error("Received malformed message without ephemeral key")
+				metrics.RecordHolePunchEvent(relayIfname, "error")
 				// Return the buffer to the pool for reuse and continue with next packet
 				bufferPool.Put(packet.data[:1500])
 				continue
@@ -402,6 +412,7 @@ func (s *UDPProxyServer) packetWorker() {
 			decryptedData, err := s.decryptMessage(encMsg)
 			if err != nil {
 				// logger.Error("Failed to decrypt message: %v", err)
+				metrics.RecordHolePunchEvent(relayIfname, "error")
 				// Return the buffer to the pool for reuse and continue with next packet
 				bufferPool.Put(packet.data[:1500])
 				continue
@@ -411,6 +422,7 @@ func (s *UDPProxyServer) packetWorker() {
 			var msg HolePunchMessage
 			if err := json.Unmarshal(decryptedData, &msg); err != nil {
 				logger.Error("Error unmarshaling decrypted message: %v", err)
+				metrics.RecordHolePunchEvent(relayIfname, "error")
 				// Return the buffer to the pool for reuse and continue with next packet
 				bufferPool.Put(packet.data[:1500])
 				continue
@@ -420,22 +432,75 @@ func (s *UDPProxyServer) packetWorker() {
 				NewtID:            msg.NewtID,
 				OlmID:             msg.OlmID,
 				Token:             msg.Token,
-				IP:                effectiveAddr.IP.String(),
-				Port:              effectiveAddr.Port,
+				IP:                packet.remoteAddr.IP.String(),
+				Port:              packet.remoteAddr.Port,
 				Timestamp:         time.Now().Unix(),
 				ReachableAt:       s.ReachableAt,
 				ExitNodePublicKey: s.privateKey.PublicKey().String(),
 				ClientPublicKey:   msg.PublicKey,
 			}
-			logger.Debug("Created endpoint from packet remoteAddr %s: IP=%s, Port=%d", effectiveAddr.String(), endpoint.IP, endpoint.Port)
-			s.notifyServer(endpoint)
+			logger.Debug("Created endpoint from packet remoteAddr %s: IP=%s, Port=%d", packet.remoteAddr.String(), endpoint.IP, endpoint.Port)
+
+			// Check if anything meaningful changed before queuing an HTTP notification.
+			// The cache expires after 2.5 s so the server always receives a fresh
+			// timestamp within its 5-second staleness window.
+			const endpointCacheTTL = 2500 * time.Millisecond
+			cacheKey := endpoint.OlmID + ":" + endpoint.NewtID
+			newState := cachedEndpointState{
+				OlmID:     endpoint.OlmID,
+				NewtID:    endpoint.NewtID,
+				Token:     endpoint.Token,
+				IP:        endpoint.IP,
+				Port:      endpoint.Port,
+				PublicKey: endpoint.ClientPublicKey,
+			}
+			if cached, ok := s.lastEndpointCache.Load(cacheKey); ok {
+				entry := cached.(cachedEndpointEntry)
+				if entry.state == newState && time.Since(entry.cachedAt) < endpointCacheTTL {
+					// Endpoint unchanged and cache still fresh - skip the HTTP call.
+					logger.Debug("Endpoint unchanged for %s, skipping notification", cacheKey)
+					metrics.RecordHolePunchEvent(relayIfname, "deduplicated")
+					s.clearSessionsForIP(endpoint.IP)
+					metrics.RecordHolePunchEvent(relayIfname, "success")
+					bufferPool.Put(packet.data[:1500])
+					continue
+				}
+			}
+			s.lastEndpointCache.Store(cacheKey, cachedEndpointEntry{state: newState, cachedAt: time.Now()})
+
+			// Queue the notification asynchronously so the hot path is not blocked by HTTP.
+			select {
+			case s.notifyChan <- endpoint:
+			case <-s.ctx.Done():
+				// shutting down
+			default:
+				logger.Debug("Notification queue full, dropping hole punch notification for %s:%d", endpoint.IP, endpoint.Port)
+				metrics.RecordHolePunchEvent(relayIfname, "queue_full")
+			}
 			s.clearSessionsForIP(endpoint.IP) // Clear sessions for this IP to allow re-establishment
+			metrics.RecordHolePunchEvent(relayIfname, "success")
 		}
 		// Return the buffer to the pool for reuse.
 		bufferPool.Put(packet.data[:1500])
 	}
 }
 
+// endpointNotifierWorker drains the notifyChan and performs the HTTP notification for each
+// hole-punch endpoint. Running several of these keeps latency low even when the server is slow.
+func (s *UDPProxyServer) endpointNotifierWorker() {
+	for {
+		select {
+		case endpoint, ok := <-s.notifyChan:
+			if !ok {
+				return
+			}
+			s.notifyServer(endpoint)
+		case <-s.ctx.Done():
+			return
+		}
+	}
+}
+
 // decryptMessage decrypts the message using the server's private key
 func (s *UDPProxyServer) decryptMessage(encMsg EncryptedHolePunchMessage) ([]byte, error) {
 	// Parse the ephemeral public key
@@ -471,6 +536,7 @@ func (s *UDPProxyServer) decryptMessage(encMsg EncryptedHolePunchMessage) ([]byt
 }
 
 func (s *UDPProxyServer) fetchInitialMappings() error {
+	logger.Info("Requesting initial proxy mappings")
 	body := bytes.NewBuffer([]byte(fmt.Sprintf(`{"publicKey": "%s"}`, s.privateKey.PublicKey().String())))
 	resp, err := http.Post(s.serverURL+"/gerbil/get-all-relays", "application/json", body)
 	if err != nil {
@@ -482,22 +548,82 @@ func (s *UDPProxyServer) fetchInitialMappings() error {
 		return fmt.Errorf("server returned non-OK status: %d, body: %s",
 			resp.StatusCode, string(body))
 	}
-	data, err := io.ReadAll(resp.Body)
+	logger.Info("Received initial mappings, streaming decode")
+
+	// Stream-decode the response instead of buffering the entire body
+	// (which can be 100MB+) and then re-walking it with json.Unmarshal.
+	// This both lowers peak memory and lets us start populating the
+	// sync.Map as entries arrive.
+	dec := json.NewDecoder(bufio.NewReaderSize(resp.Body, 1<<20))
+
+	// Expect opening '{' of the top-level object.
+	tok, err := dec.Token()
 	if err != nil {
-		return fmt.Errorf("failed to read response body: %v", err)
+		return fmt.Errorf("failed to read opening token: %v", err)
 	}
-	logger.Info("Received initial mappings: %s", string(data))
-	var initialMappings InitialMappings
-	if err := json.Unmarshal(data, &initialMappings); err != nil {
-		return fmt.Errorf("failed to unmarshal initial mappings: %v", err)
+	if d, ok := tok.(json.Delim); !ok || d != '{' {
+		return fmt.Errorf("expected '{' at top level, got %v", tok)
 	}
-	// Store mappings in our sync.Map.
-	for key, mapping := range initialMappings.Mappings {
-		// Initialize LastUsed timestamp for initial mappings
-		mapping.LastUsed = time.Now()
-		s.proxyMappings.Store(key, mapping)
+
+	count := 0
+	now := time.Now()
+
+	for dec.More() {
+		keyTok, err := dec.Token()
+		if err != nil {
+			return fmt.Errorf("failed to read top-level key: %v", err)
+		}
+		key, ok := keyTok.(string)
+		if !ok {
+			return fmt.Errorf("expected string key at top level, got %T", keyTok)
+		}
+
+		if key != "mappings" {
+			// Skip unknown top-level fields without materializing them.
+			var skip json.RawMessage
+			if err := dec.Decode(&skip); err != nil {
+				return fmt.Errorf("failed to skip field %q: %v", key, err)
+			}
+			continue
+		}
+
+		// Expect opening '{' of the mappings object.
+		tok, err := dec.Token()
+		if err != nil {
+			return fmt.Errorf("failed to read mappings open: %v", err)
+		}
+		if d, ok := tok.(json.Delim); !ok || d != '{' {
+			return fmt.Errorf("expected '{' for mappings, got %v", tok)
+		}
+
+		for dec.More() {
+			mapKeyTok, err := dec.Token()
+			if err != nil {
+				return fmt.Errorf("failed to read mapping key: %v", err)
+			}
+			mapKey, ok := mapKeyTok.(string)
+			if !ok {
+				return fmt.Errorf("expected string mapping key, got %T", mapKeyTok)
+			}
+
+			var mapping ProxyMapping
+			if err := dec.Decode(&mapping); err != nil {
+				return fmt.Errorf("failed to decode mapping %q: %v", mapKey, err)
+			}
+			mapping.LastUsed = now
+			s.proxyMappings.Store(mapKey, mapping)
+			count++
+		}
+
+		// Consume closing '}' of mappings object.
+		if _, err := dec.Token(); err != nil {
+			return fmt.Errorf("failed to read mappings close: %v", err)
+		}
 	}
-	logger.Info("Loaded %d initial proxy mappings", len(initialMappings.Mappings))
+
+	metrics.RecordProxyInitialMappings(relayIfname, int64(count))
+	metrics.RecordProxyMapping(relayIfname, int64(count))
+	logger.Info("Loaded %d initial proxy mappings", count)
 	return nil
 }
 
@@ -612,7 +738,11 @@ func (s *UDPProxyServer) handleWireGuardPacket(packet []byte, remoteAddr *net.UD
 			_, err = conn.Write(packet)
 			if err != nil {
 				logger.Debug("Failed to forward handshake initiation: %v", err)
+				metrics.RecordProxyConnectionError(relayIfname, "write_udp")
+				continue
 			}
+			metrics.RecordUDPPacket(relayIfname, "wireguard", "out")
+			metrics.RecordUDPPacketSize(relayIfname, "wireguard", float64(len(packet)))
 		}
 
 	case WireGuardMessageTypeHandshakeResponse:
@@ -624,12 +754,19 @@ func (s *UDPProxyServer) handleWireGuardPacket(packet []byte, remoteAddr *net.UD
 		sessionKey := fmt.Sprintf("%d:%d", receiverIndex, senderIndex)
 
 		// Store the session information
-		s.wgSessions.Store(sessionKey, &WireGuardSession{
+		session := &WireGuardSession{
 			ReceiverIndex: receiverIndex,
 			SenderIndex:   senderIndex,
 			DestAddr:      remoteAddr,
 			LastSeen:      time.Now(),
-		})
+		}
+		if _, loaded := s.wgSessions.LoadOrStore(sessionKey, session); loaded {
+			s.wgSessions.Store(sessionKey, session)
+		} else {
+			metrics.RecordSession(relayIfname, 1)
+		}
+		// Also index by sender index for O(1) lookup in transport data path
+		s.sessionsByReceiverIndex.Store(senderIndex, session)
 
 		// Forward the response to the original sender
 		for _, dest := range proxyMapping.Destinations {
@@ -648,28 +785,26 @@ func (s *UDPProxyServer) handleWireGuardPacket(packet []byte, remoteAddr *net.UD
 			_, err = conn.Write(packet)
 			if err != nil {
 				logger.Error("Failed to forward handshake response: %v", err)
+				metrics.RecordProxyConnectionError(relayIfname, "write_udp")
+				continue
 			}
+			metrics.RecordUDPPacket(relayIfname, "wireguard", "out")
+			metrics.RecordUDPPacketSize(relayIfname, "wireguard", float64(len(packet)))
 		}
 
 	case WireGuardMessageTypeTransportData:
 		// Data packet: forward only to the established session peer
 		// logger.Debug("Received transport data with receiver index %d from %s", receiverIndex, remoteAddr)
 
-		// Look up the session based on the receiver index
+		// Look up the session based on the receiver index - O(1) lookup instead of O(n) Range
 		var destAddr *net.UDPAddr
 
-		// First check for existing sessions to see if we know where to send this packet
-		s.wgSessions.Range(func(k, v interface{}) bool {
-			session := v.(*WireGuardSession)
-			// Check if session matches (read lock for check)
-			if session.GetSenderIndex() == receiverIndex {
-				// Found matching session - get dest addr and update last seen
-				destAddr = session.GetDestAddr()
-				session.UpdateLastSeen()
-				return false // stop iteration
-			}
-			return true // continue iteration
-		})
+		// Fast path: direct index lookup by receiver index
+		if sessionObj, ok := s.sessionsByReceiverIndex.Load(receiverIndex); ok {
+			session := sessionObj.(*WireGuardSession)
+			destAddr = session.GetDestAddr()
+			session.UpdateLastSeen()
+		}
 
 		if destAddr != nil {
 			// We found a specific peer to forward to
@@ -685,7 +820,11 @@ func (s *UDPProxyServer) handleWireGuardPacket(packet []byte, remoteAddr *net.UD
 			_, err = conn.Write(packet)
 			if err != nil {
 				logger.Debug("Failed to forward transport data: %v", err)
+				metrics.RecordProxyConnectionError(relayIfname, "write_udp")
+				return
 			}
+			metrics.RecordUDPPacket(relayIfname, "wireguard", "out")
+			metrics.RecordUDPPacketSize(relayIfname, "wireguard", float64(len(packet)))
 		} else {
 			// No known session, fall back to forwarding to all peers
 			logger.Debug("No session found for receiver index %d, forwarding to all destinations", receiverIndex)
@@ -708,7 +847,11 @@ func (s *UDPProxyServer) handleWireGuardPacket(packet []byte, remoteAddr *net.UD
 				_, err = conn.Write(packet)
 				if err != nil {
 					logger.Debug("Failed to forward transport data: %v", err)
+					metrics.RecordProxyConnectionError(relayIfname, "write_udp")
+					continue
 				}
+				metrics.RecordUDPPacket(relayIfname, "wireguard", "out")
+				metrics.RecordUDPPacketSize(relayIfname, "wireguard", float64(len(packet)))
 			}
 		}
 
@@ -733,7 +876,11 @@ func (s *UDPProxyServer) handleWireGuardPacket(packet []byte, remoteAddr *net.UD
 			_, err = conn.Write(packet)
 			if err != nil {
 				logger.Error("Failed to forward WireGuard packet: %v", err)
+				metrics.RecordProxyConnectionError(relayIfname, "write_udp")
+				continue
 			}
+			metrics.RecordUDPPacket(relayIfname, "wireguard", "out")
+			metrics.RecordUDPPacketSize(relayIfname, "wireguard", float64(len(packet)))
 		}
 	}
 }
@@ -744,21 +891,35 @@ func (s *UDPProxyServer) getOrCreateConnection(destAddr *net.UDPAddr, remoteAddr
 	// Check if we have an existing connection
 	if conn, ok := s.connections.Load(key); ok {
 		destConn := conn.(*DestinationConn)
-		destConn.lastUsed = time.Now()
+		destConn.lastUsed.Store(time.Now().UnixNano())
 		return destConn.conn, nil
 	}
 
+	// Enforce a hard cap on concurrent sockets so a burst of peer churn can't
+	// exhaust the host's ephemeral ports/fds. Evict the least-recently-used
+	// connection to make room instead of growing unbounded.
+	if s.connectionCount.Load() >= maxUDPConnections {
+		s.evictLRUConnection()
+	}
+
 	// Create new connection
 	newConn, err := net.DialUDP("udp", nil, destAddr)
 	if err != nil {
+		metrics.RecordProxyConnectionError(relayIfname, "dial_udp")
 		return nil, fmt.Errorf("failed to create UDP connection: %v", err)
 	}
 
-	// Store the new connection
-	s.connections.Store(key, &DestinationConn{
-		conn:     newConn,
-		lastUsed: time.Now(),
-	})
+	destConn := &DestinationConn{conn: newConn}
+	destConn.lastUsed.Store(time.Now().UnixNano())
+
+	// Store the new connection. If another goroutine raced us and already
+	// created one for this key, close ours and use theirs instead.
+	if existing, loaded := s.connections.LoadOrStore(key, destConn); loaded {
+		newConn.Close()
+		return existing.(*DestinationConn).conn, nil
+	}
+	s.connectionCount.Add(1)
+	metrics.RecordUDPConnection(relayIfname, 1)
 
 	// Start a goroutine to handle responses
 	go s.handleResponses(newConn, destAddr, remoteAddr)
@@ -766,6 +927,33 @@ func (s *UDPProxyServer) getOrCreateConnection(destAddr *net.UDPAddr, remoteAddr
 	return newConn, nil
 }
 
+// evictLRUConnection closes and removes the least-recently-used destination
+// connection so a new one can be created under the concurrent connection cap.
+func (s *UDPProxyServer) evictLRUConnection() {
+	var oldestKey interface{}
+	var oldestConn *DestinationConn
+	var oldestTime int64
+
+	s.connections.Range(func(key, value interface{}) bool {
+		destConn := value.(*DestinationConn)
+		lu := destConn.lastUsed.Load()
+		if oldestKey == nil || lu < oldestTime {
+			oldestKey = key
+			oldestConn = destConn
+			oldestTime = lu
+		}
+		return true
+	})
+
+	if oldestKey != nil {
+		s.connections.Delete(oldestKey)
+		oldestConn.conn.Close()
+		s.connectionCount.Add(-1)
+		metrics.RecordUDPConnection(relayIfname, -1)
+		metrics.RecordProxyCleanupRemoved(relayIfname, "conn_evicted", 1)
+	}
+}
+
 func (s *UDPProxyServer) handleResponses(conn *net.UDPConn, destAddr *net.UDPAddr, remoteAddr *net.UDPAddr) {
 	buffer := make([]byte, 1500)
 	for {
@@ -774,6 +962,8 @@ func (s *UDPProxyServer) handleResponses(conn *net.UDPConn, destAddr *net.UDPAdd
 			logger.Debug("Error reading response from %s: %v", destAddr.String(), err)
 			return
 		}
+		metrics.RecordUDPPacket(relayIfname, "wireguard", "in")
+		metrics.RecordUDPPacketSize(relayIfname, "wireguard", float64(n))
 
 		// Process the response to track sessions if it's a WireGuard packet
 		if n > 0 && buffer[0] >= 1 && buffer[0] <= 4 {
@@ -781,12 +971,19 @@ func (s *UDPProxyServer) handleResponses(conn *net.UDPConn, destAddr *net.UDPAdd
 			if ok && buffer[0] == WireGuardMessageTypeHandshakeResponse {
 				// Store the session mapping for the handshake response
 				sessionKey := fmt.Sprintf("%d:%d", senderIndex, receiverIndex)
-				s.wgSessions.Store(sessionKey, &WireGuardSession{
+				session := &WireGuardSession{
 					ReceiverIndex: receiverIndex,
 					SenderIndex:   senderIndex,
 					DestAddr:      destAddr,
 					LastSeen:      time.Now(),
-				})
+				}
+				if _, loaded := s.wgSessions.LoadOrStore(sessionKey, session); loaded {
+					s.wgSessions.Store(sessionKey, session)
+				} else {
+					metrics.RecordSession(relayIfname, 1)
+				}
+				// Also index by sender index for O(1) lookup
+				s.sessionsByReceiverIndex.Store(senderIndex, session)
 				logger.Debug("Stored session mapping: %s -> %s", sessionKey, destAddr.String())
 			} else if ok && buffer[0] == WireGuardMessageTypeTransportData {
 				// Track communication pattern for session rebuilding (reverse direction)
@@ -798,26 +995,43 @@ func (s *UDPProxyServer) handleResponses(conn *net.UDPConn, destAddr *net.UDPAdd
 		_, err = s.conn.WriteToUDP(buffer[:n], remoteAddr)
 		if err != nil {
 			logger.Error("Failed to forward response: %v", err)
+			metrics.RecordProxyConnectionError(relayIfname, "write_udp")
+			continue
 		}
+		metrics.RecordUDPPacket(relayIfname, "wireguard", "out")
+		metrics.RecordUDPPacketSize(relayIfname, "wireguard", float64(n))
 	}
 }
 
 // Add a cleanup method to periodically remove idle connections
 func (s *UDPProxyServer) cleanupIdleConnections() {
-	ticker := time.NewTicker(5 * time.Minute)
+	// Ticker interval and idle threshold were previously 5min/10min, meaning
+	// a socket could sit open for up to 15 minutes after going idle. Under a
+	// reconnect/churn burst that lag is enough to exhaust ephemeral ports
+	// before cleanup catches up, so both are tightened here.
+	ticker := time.NewTicker(1 * time.Minute)
 	defer ticker.Stop()
 	for {
 		select {
 		case <-ticker.C:
-			now := time.Now()
+			cleanupStart := time.Now()
+			now := time.Now().UnixNano()
+			removed := int64(0)
 			s.connections.Range(func(key, value interface{}) bool {
 				destConn := value.(*DestinationConn)
-				if now.Sub(destConn.lastUsed) > 10*time.Minute {
+				if now-destConn.lastUsed.Load() > int64(5*time.Minute) {
 					destConn.conn.Close()
 					s.connections.Delete(key)
+					removed++
 				}
 				return true
 			})
+			if removed > 0 {
+				s.connectionCount.Add(-removed)
+				metrics.RecordUDPConnection(relayIfname, -removed)
+				metrics.RecordProxyCleanupRemoved(relayIfname, "conn", removed)
+			}
+			metrics.RecordProxyIdleCleanupDuration(relayIfname, "conn", time.Since(cleanupStart).Seconds())
 		case <-s.ctx.Done():
 			return
 		}
@@ -832,16 +1046,20 @@ func (s *UDPProxyServer) cleanupIdleSessions() {
 	for {
 		select {
 		case <-ticker.C:
+			cleanupStart := time.Now()
 			now := time.Now()
 			s.wgSessions.Range(func(key, value interface{}) bool {
 				session := value.(*WireGuardSession)
 				// Use thread-safe method to read LastSeen
 				if now.Sub(session.GetLastSeen()) > 15*time.Minute {
 					s.wgSessions.Delete(key)
+					metrics.RecordSession(relayIfname, -1)
+					metrics.RecordProxyCleanupRemoved(relayIfname, "session", 1)
 					logger.Debug("Removed idle session: %s", key)
 				}
 				return true
 			})
+			metrics.RecordProxyIdleCleanupDuration(relayIfname, "session", time.Since(cleanupStart).Seconds())
 		case <-s.ctx.Done():
 			return
 		}
@@ -855,16 +1073,20 @@ func (s *UDPProxyServer) cleanupIdleProxyMappings() {
 	for {
 		select {
 		case <-ticker.C:
+			cleanupStart := time.Now()
 			now := time.Now()
 			s.proxyMappings.Range(func(key, value interface{}) bool {
 				mapping := value.(ProxyMapping)
 				// Remove mappings that haven't been used in 30 minutes
 				if now.Sub(mapping.LastUsed) > 30*time.Minute {
 					s.proxyMappings.Delete(key)
+					metrics.RecordProxyMapping(relayIfname, -1)
+					metrics.RecordProxyCleanupRemoved(relayIfname, "proxy_mapping", 1)
 					logger.Debug("Removed idle proxy mapping: %s", key)
 				}
 				return true
 			})
+			metrics.RecordProxyIdleCleanupDuration(relayIfname, "proxy_mapping", time.Since(cleanupStart).Seconds())
 		case <-s.ctx.Done():
 			return
 		}
@@ -907,6 +1129,11 @@ func (s *UDPProxyServer) notifyServer(endpoint ClientEndpoint) {
 	key := fmt.Sprintf("%s:%d", endpoint.IP, endpoint.Port)
 	logger.Debug("About to store proxy mapping with key: %s (from endpoint IP=%s, Port=%d)", key, endpoint.IP, endpoint.Port)
 	mapping.LastUsed = time.Now()
+	if _, existed := s.proxyMappings.Load(key); existed {
+		metrics.RecordProxyMappingUpdate(relayIfname)
+	} else {
+		metrics.RecordProxyMapping(relayIfname, 1)
+	}
 	s.proxyMappings.Store(key, mapping)
 
 	logger.Debug("Stored proxy mapping for %s with %d destinations (timestamp: %v)", key, len(mapping.Destinations), mapping.LastUsed)
@@ -919,6 +1146,11 @@ func (s *UDPProxyServer) UpdateProxyMapping(sourceIP string, sourcePort int, des
 		Destinations: destinations,
 		LastUsed:     time.Now(),
 	}
+	if _, existed := s.proxyMappings.Load(key); existed {
+		metrics.RecordProxyMappingUpdate(relayIfname)
+	} else {
+		metrics.RecordProxyMapping(relayIfname, 1)
+	}
 	s.proxyMappings.Store(key, mapping)
 }
 
@@ -960,6 +1192,10 @@ func (s *UDPProxyServer) clearConnectionsForWGIP(wgIP string) {
 	for _, key := range keysToDelete {
 		s.connections.Delete(key)
 	}
+	if len(keysToDelete) > 0 {
+		s.connectionCount.Add(-int64(len(keysToDelete)))
+		metrics.RecordUDPConnection(relayIfname, -int64(len(keysToDelete)))
+	}
 
 	logger.Info("Cleared %d connections for WG IP: %s", len(keysToDelete), wgIP)
 }
@@ -985,6 +1221,10 @@ func (s *UDPProxyServer) clearSessionsForIP(ip string) {
 	for _, key := range keysToDelete {
 		s.wgSessions.Delete(key)
 	}
+	if len(keysToDelete) > 0 {
+		metrics.RecordSession(relayIfname, -int64(len(keysToDelete)))
+		metrics.RecordProxyCleanupRemoved(relayIfname, "session", int64(len(keysToDelete)))
+	}
 
 	logger.Debug("Cleared %d sessions for WG IP: %s", len(keysToDelete), ip)
 }
@@ -1145,35 +1385,64 @@ func (s *UDPProxyServer) trackCommunicationPattern(fromAddr, toAddr *net.UDPAddr
 			pattern.LastFromDest = now
 		}
 
-		s.commPatterns.Store(patternKey, pattern)
+		if _, loaded := s.commPatterns.LoadOrStore(patternKey, pattern); !loaded {
+			metrics.RecordCommPattern(relayIfname, 1)
+		}
 	}
 }
 
 // tryRebuildSession attempts to rebuild a WireGuard session from communication patterns
 func (s *UDPProxyServer) tryRebuildSession(pattern *CommunicationPattern) {
+	// Require both indices and a minimum amount of bidirectional traffic
+	if pattern.ClientIndex == 0 || pattern.DestIndex == 0 || pattern.PacketCount < 4 {
+		return
+	}
+
 	// Check if we have bidirectional communication within a reasonable time window
 	timeDiff := pattern.LastFromClient.Sub(pattern.LastFromDest)
 	if timeDiff < 0 {
 		timeDiff = -timeDiff
 	}
+	if timeDiff >= 30*time.Second {
+		return
+	}
 
-	// Only rebuild if we have recent bidirectional communication and both indices
-	if timeDiff < 30*time.Second && pattern.ClientIndex != 0 && pattern.DestIndex != 0 && pattern.PacketCount >= 4 {
-		// Create session mapping: client's index maps to destination
-		sessionKey := fmt.Sprintf("%d:%d", pattern.DestIndex, pattern.ClientIndex)
+	sessionKey := fmt.Sprintf("%d:%d", pattern.DestIndex, pattern.ClientIndex)
+	destStr := pattern.ToDestination.String()
 
-		// Check if we already have this session
-		if _, exists := s.wgSessions.Load(sessionKey); !exists {
-			s.wgSessions.Store(sessionKey, &WireGuardSession{
-				ReceiverIndex: pattern.DestIndex,
-				SenderIndex:   pattern.ClientIndex,
-				DestAddr:      pattern.ToDestination,
-				LastSeen:      time.Now(),
-			})
-			logger.Info("Rebuilt WireGuard session from communication pattern: %s -> %s (packets: %d)",
-				sessionKey, pattern.ToDestination.String(), pattern.PacketCount)
+	// Fast path: if a matching session already exists, just refresh LastSeen and bail out.
+	// This prevents log spam and repeated work for every packet of an established flow.
+	if existing, ok := s.wgSessions.Load(sessionKey); ok {
+		sess := existing.(*WireGuardSession)
+		if da := sess.GetDestAddr(); da != nil && da.String() == destStr {
+			sess.UpdateLastSeen()
+			// Make sure the receiver-index fast-path is populated so future packets
+			// don't keep falling back to broadcast + pattern tracking.
+			if _, indexed := s.sessionsByReceiverIndex.Load(pattern.ClientIndex); !indexed {
+				s.sessionsByReceiverIndex.Store(pattern.ClientIndex, sess)
+			}
+			return
 		}
 	}
+
+	// Create or replace the session mapping
+	session := &WireGuardSession{
+		ReceiverIndex: pattern.DestIndex,
+		SenderIndex:   pattern.ClientIndex,
+		DestAddr:      pattern.ToDestination,
+		LastSeen:      time.Now(),
+	}
+	if _, loaded := s.wgSessions.LoadOrStore(sessionKey, session); loaded {
+		s.wgSessions.Store(sessionKey, session)
+	} else {
+		metrics.RecordSession(relayIfname, 1)
+		metrics.RecordSessionRebuilt(relayIfname)
+	}
+	// Index by client receiver index so the transport-data fast path can find it.
+	s.sessionsByReceiverIndex.Store(pattern.ClientIndex, session)
+
+	logger.Info("Rebuilt WireGuard session from communication pattern: %s -> %s (packets: %d)",
+		sessionKey, destStr, pattern.PacketCount)
 }
 
 // cleanupIdleCommunicationPatterns periodically removes idle communication patterns
@@ -1207,6 +1476,7 @@ func (s *UDPProxyServer) cleanupIdleCommunicationPatterns() {
 	for {
 		select {
 		case <-ticker.C:
+			cleanupStart := time.Now()
 			now := time.Now()
 			s.commPatterns.Range(func(key, value interface{}) bool {
 				pattern := value.(*CommunicationPattern)
@@ -1220,10 +1490,13 @@ func (s *UDPProxyServer) cleanupIdleCommunicationPatterns() {
 				// Remove patterns that haven't had activity in 20 minutes
 				if now.Sub(lastActivity) > 20*time.Minute {
 					s.commPatterns.Delete(key)
+					metrics.RecordCommPattern(relayIfname, -1)
+					metrics.RecordProxyCleanupRemoved(relayIfname, "comm_pattern", 1)
 					logger.Debug("Removed idle communication pattern: %s", key)
 				}
 				return true
 			})
+			metrics.RecordProxyIdleCleanupDuration(relayIfname, "comm_pattern", time.Since(cleanupStart).Seconds())
 		case <-s.ctx.Done():
 			return
 		}