feat: Add per-instance OTEL resource attributes for metrics and logs

CLAUDE.md

@@ -210,6 +210,8 @@ See README.md for complete configuration reference.
 - Prometheus metrics exported to OTLP HTTP endpoint
 - Tracks: active sessions, audio bytes, transcription latency, backend errors
 - Only enabled when `OTLP_ENDPOINT` is set
+- Each container instance is differentiated via `CLOUDFLARE_DURABLE_OBJECT_ID` (falls back to random UUID for local dev). Metrics use `service.instance.id` (standard OTEL, Mimir-friendly). Logs use `runId` (custom name to avoid Loki auto-indexing it as a high-cardinality label).
+- Container location is tagged via `city` (from `CLOUDFLARE_LOCATION`) and `country` (from `CLOUDFLARE_COUNTRY_A2`)
 
 **Logging** (`src/logger.ts`)
 - Uses Winston with OTLP logs transport

src/telemetry.ts

@@ -28,21 +28,57 @@ let loggerProvider: LoggerProvider | null = null;
 let meter: Meter | null = null;
 
 /**
- * Create shared resource with common attributes for both metrics and logs.
+ * Build the base resource attributes shared by metrics and logs.
  */
-function createResource(): Resource {
-	const resourceAttributes: Record<string, string> = {
+function baseResourceAttributes(): Record<string, string> {
+	const attrs: Record<string, string> = {
 		[ATTR_SERVICE_NAME]: 'opus-transcriber-proxy',
 	};
 
 	if (config.otlp.env) {
-		resourceAttributes[ATTR_DEPLOYMENT_ENVIRONMENT] = config.otlp.env;
-		resourceAttributes['env'] = config.otlp.env;
+		attrs[ATTR_DEPLOYMENT_ENVIRONMENT] = config.otlp.env;
+		attrs['env'] = config.otlp.env;
 	}
 
-	Object.assign(resourceAttributes, config.otlp.resourceAttributes);
+	// Cloudflare container location context (injected automatically by CF runtime)
+	if (process.env.CLOUDFLARE_LOCATION) {
+		attrs['city'] = process.env.CLOUDFLARE_LOCATION;
+	}
+	if (process.env.CLOUDFLARE_COUNTRY_A2) {
+		attrs['country'] = process.env.CLOUDFLARE_COUNTRY_A2;
+	}
+
+	Object.assign(attrs, config.otlp.resourceAttributes);
+
+	return attrs;
+}
+
+// Instance ID generated once per process — shared between metrics and logs resources.
+// CONTAINER_INSTANCE_NAME is the human-readable name passed to getContainer() by the worker
+// (equals the session/meeting ID in session routing mode). Falls back to DO hex hash or random UUID.
+const instanceId = process.env.CONTAINER_INSTANCE_NAME || process.env.CLOUDFLARE_DURABLE_OBJECT_ID || crypto.randomUUID();
 
-	return new Resource(resourceAttributes);
+/**
+ * Resource for metrics (Mimir). Uses the standard 'service.instance.id' attribute
+ * so each container produces unique series for correct aggregation.
+ */
+function createMetricsResource(): Resource {
+	const attrs = baseResourceAttributes();
+	attrs['service.instance.id'] = instanceId;
+	return new Resource(attrs);
+}
+
+/**
+ * Resource for logs (Loki). Sets a static 'service.instance.id' to override the OTEL SDK's
+ * auto-detected value. This prevents Loki from creating a unique indexed stream per container
+ * (high-cardinality explosion). The per-container identity is stored in 'runId' instead — a
+ * name chosen to avoid Loki auto-indexing (unlike 'session_id' or 'container_id').
+ */
+function createLogsResource(): Resource {
+	const attrs = baseResourceAttributes();
+	attrs['service.instance.id'] = 'opus-transcriber-proxy';
+	attrs['runId'] = instanceId;
+	return new Resource(attrs);
 }
 
 /**
@@ -57,7 +93,7 @@ export function initTelemetry(): void {
 
 	logger.info(`Initializing OpenTelemetry metrics, endpoint=${config.otlp.endpoint}`);
 
-	const resource = createResource();
+	const resource = createMetricsResource();
 
 	// Create OTLP exporter
 	const exporterConfig: { url: string; headers?: Record<string, string> } = {
@@ -137,7 +173,7 @@ export function initTelemetryLogs(): void {
 	// Use console.log to avoid circular dependency with logger
 	console.log(`Initializing OpenTelemetry logs, endpoint=${config.otlp.endpoint}/v1/logs`);
 
-	const resource = createResource();
+	const resource = createLogsResource();
 
 	const logExporterConfig: { url: string; headers?: Record<string, string> } = {
 		url: `${config.otlp.endpoint}/v1/logs`,

worker/index.ts

@@ -35,6 +35,41 @@ interface TranscriptionMessage {
 	language?: string;
 }
 
+/**
+ * Build the container env vars from the worker env bindings.
+ * Shared between the class-level default and startAndWaitForPorts() override
+ * (which adds CONTAINER_INSTANCE_NAME — not available at class init time).
+ */
+function buildContainerEnvVars(env: Env): Record<string, string> {
+	return {
+		OPENAI_API_KEY: env.OPENAI_API_KEY,
+		OPENAI_MODEL: env.OPENAI_MODEL || 'gpt-4o-transcribe',
+		GEMINI_API_KEY: env.GEMINI_API_KEY || '',
+		DEEPGRAM_API_KEY: env.DEEPGRAM_API_KEY || '',
+		DEEPGRAM_MODEL: env.DEEPGRAM_MODEL || 'nova-3-general',
+		DEEPGRAM_DETECT_LANGUAGE: env.DEEPGRAM_DETECT_LANGUAGE || 'true',
+		DEEPGRAM_INCLUDE_LANGUAGE: env.DEEPGRAM_INCLUDE_LANGUAGE || 'false',
+		DEEPGRAM_PUNCTUATE: env.DEEPGRAM_PUNCTUATE || 'true',
+		DEEPGRAM_ENCODING: env.DEEPGRAM_ENCODING || 'opus',
+		DEEPGRAM_TAGS: env.DEEPGRAM_TAGS || '',
+		PROVIDERS_PRIORITY: env.PROVIDERS_PRIORITY || 'openai',
+		FORCE_COMMIT_TIMEOUT: env.FORCE_COMMIT_TIMEOUT || '2',
+		DEBUG: env.DEBUG || 'true',
+		ROUTING_MODE: env.ROUTING_MODE || 'session',
+		CONTAINER_POOL_SIZE: env.CONTAINER_POOL_SIZE || '5',
+		MAX_CONNECTIONS_PER_CONTAINER: env.MAX_CONNECTIONS_PER_CONTAINER || '10',
+		MIN_CONTAINERS: env.MIN_CONTAINERS || '2',
+		SCALE_DOWN_IDLE_TIME: env.SCALE_DOWN_IDLE_TIME || '600000',
+		TRANSLATION_MIXING_MODE: env.TRANSLATION_MIXING_MODE || 'true',
+		OTLP_ENDPOINT: env.OTLP_ENDPOINT || '',
+		OTLP_ENV: env.OTLP_ENV || '',
+		OTLP_RESOURCE_ATTRIBUTES: env.OTLP_RESOURCE_ATTRIBUTES || '',
+		OTLP_HEADERS: env.OTLP_HEADERS || '',
+		PORT: '8080',
+		HOST: '0.0.0.0',
+	};
+}
+
 /**
  * TranscriberContainer wraps the Node.js transcription server
  * and forwards WebSocket requests to it.
@@ -52,35 +87,8 @@ export class TranscriberContainer extends Container<Env> {
 	// For session-based routing: Keep this shorter (containers are session-specific)
 	sleepAfter = this.env.SLEEP_AFTER || '1m';
 
-	// Pass environment variables to the container
-	envVars: Record<string, string> = {
-		// These will be available as process.env in the container
-		OPENAI_API_KEY: this.env.OPENAI_API_KEY,
-		OPENAI_MODEL: this.env.OPENAI_MODEL || 'gpt-4o-transcribe',
-		GEMINI_API_KEY: this.env.GEMINI_API_KEY || '',
-		DEEPGRAM_API_KEY: this.env.DEEPGRAM_API_KEY || '',
-		DEEPGRAM_MODEL: this.env.DEEPGRAM_MODEL || 'nova-3-general',
-		DEEPGRAM_DETECT_LANGUAGE: this.env.DEEPGRAM_DETECT_LANGUAGE || 'true',
-		DEEPGRAM_INCLUDE_LANGUAGE: this.env.DEEPGRAM_INCLUDE_LANGUAGE || 'false',
-		DEEPGRAM_PUNCTUATE: this.env.DEEPGRAM_PUNCTUATE || 'true',
-		DEEPGRAM_ENCODING: this.env.DEEPGRAM_ENCODING || 'opus',
-		DEEPGRAM_TAGS: this.env.DEEPGRAM_TAGS || '',
-		PROVIDERS_PRIORITY: this.env.PROVIDERS_PRIORITY || 'openai',
-		FORCE_COMMIT_TIMEOUT: this.env.FORCE_COMMIT_TIMEOUT || '2',
-		DEBUG: this.env.DEBUG || 'true',
-		ROUTING_MODE: this.env.ROUTING_MODE || 'session',
-		CONTAINER_POOL_SIZE: this.env.CONTAINER_POOL_SIZE || '5',
-		MAX_CONNECTIONS_PER_CONTAINER: this.env.MAX_CONNECTIONS_PER_CONTAINER || '10',
-		MIN_CONTAINERS: this.env.MIN_CONTAINERS || '2',
-		SCALE_DOWN_IDLE_TIME: this.env.SCALE_DOWN_IDLE_TIME || '600000',
-		TRANSLATION_MIXING_MODE: this.env.TRANSLATION_MIXING_MODE || 'true',
-		OTLP_ENDPOINT: this.env.OTLP_ENDPOINT || '',
-		OTLP_ENV: this.env.OTLP_ENV || '',
-		OTLP_RESOURCE_ATTRIBUTES: this.env.OTLP_RESOURCE_ATTRIBUTES || '',
-		OTLP_HEADERS: this.env.OTLP_HEADERS || '',
-		PORT: '8080',
-		HOST: '0.0.0.0',
-	};
+	// Pass environment variables to the container (default, without instance name)
+	envVars: Record<string, string> = buildContainerEnvVars(this.env);
 
 	override onStart() {
 		console.log('Transcriber container started');
@@ -513,6 +521,9 @@ export default {
 		// This is required for the fetch to work properly
 		try {
 			await container.startAndWaitForPorts({
+				startOptions: {
+					envVars: { ...buildContainerEnvVars(env), CONTAINER_INSTANCE_NAME: containerInstanceId },
+				},
 				cancellationOptions: {
 					waitInterval: 100, // Poll every 100ms (default: 1000ms)
 				},