feat: improve telemetry (#2590)

# What does this PR do?

* Use a single env variable to setup OTEL endpoint
* Update telemetry provider doc
* Update general telemetry doc with the metric with generate
* Left a script to setup telemetry for testing

Closes: https://github.com/meta-llama/llama-stack/issues/783

Note to reviewer: the `setup_telemetry.sh` script was useful for me, it
was nicely generated by AI, if we don't want it in the repo, and I can
delete it, and I would understand.

Signed-off-by: Sébastien Han <seb@redhat.com>

docs/source/building_applications/telemetry.md

@@ -24,37 +24,106 @@ structured_log_event = SpanStartPayload(name="my_span", parent_span_id="parent_s
 - **Spans**: Represent operations with timing and hierarchical relationships
 - **Traces**: Collection of related spans forming a complete request flow
 
+### Metrics
+
+Llama Stack automatically generates metrics during inference operations. These metrics are aggregated at the **inference request level** and provide insights into token usage and model performance.
+
+#### Available Metrics
+
+The following metrics are automatically generated for each inference request:
+
+| Metric Name | Type | Unit | Description | Labels |
+|-------------|------|------|-------------|--------|
+| `llama_stack_prompt_tokens_total` | Counter | `tokens` | Number of tokens in the input prompt | `model_id`, `provider_id` |
+| `llama_stack_completion_tokens_total` | Counter | `tokens` | Number of tokens in the generated response | `model_id`, `provider_id` |
+| `llama_stack_tokens_total` | Counter | `tokens` | Total tokens used (prompt + completion) | `model_id`, `provider_id` |
+
+#### Metric Generation Flow
+
+1. **Token Counting**: During inference operations (chat completion, completion, etc.), the system counts tokens in both input prompts and generated responses
+2. **Metric Construction**: For each request, `MetricEvent` objects are created with the token counts
+3. **Telemetry Logging**: Metrics are sent to the configured telemetry sinks
+4. **OpenTelemetry Export**: When OpenTelemetry is enabled, metrics are exposed as standard OpenTelemetry counters
+
+#### Metric Aggregation Level
+
+All metrics are generated and aggregated at the **inference request level**. This means:
+
+- Each individual inference request generates its own set of metrics
+- Metrics are not pre-aggregated across multiple requests
+- Aggregation (sums, averages, etc.) can be performed by your observability tools (Prometheus, Grafana, etc.)
+- Each metric includes labels for `model_id` and `provider_id` to enable filtering and grouping
+
+#### Example Metric Event
+
+```python
+MetricEvent(
+    trace_id="1234567890abcdef",
+    span_id="abcdef1234567890",
+    metric="total_tokens",
+    value=150,
+    timestamp=1703123456.789,
+    unit="tokens",
+    attributes={"model_id": "meta-llama/Llama-3.2-3B-Instruct", "provider_id": "tgi"},
+)
+```
+
+#### Querying Metrics
+
+When using the OpenTelemetry sink, metrics are exposed in standard OpenTelemetry format and can be queried through:
+
+- **Prometheus**: Scrape metrics from the OpenTelemetry Collector's metrics endpoint
+- **Grafana**: Create dashboards using Prometheus as a data source
+- **OpenTelemetry Collector**: Forward metrics to other observability systems
+
+Example Prometheus queries:
+```promql
+# Total tokens used across all models
+sum(llama_stack_tokens_total)
+
+# Tokens per model
+sum by (model_id) (llama_stack_tokens_total)
+
+# Average tokens per request
+rate(llama_stack_tokens_total[5m])
+```
+
 ### Sinks
-- **OpenTelemetry**: Send events to an OpenTelemetry Collector. This is useful for visualizing traces in a tool like Jaeger.
+- **OpenTelemetry**: Send events to an OpenTelemetry Collector. This is useful for visualizing traces in a tool like Jaeger and collecting metrics for Prometheus.
 - **SQLite**: Store events in a local SQLite database. This is needed if you want to query the events later through the Llama Stack API.
 - **Console**: Print events to the console.
 
 ### Providers
 
 #### Meta-Reference Provider
-Currently, only the meta-reference provider is implemented. It can be configured to send events to three sink types:
-1) OpenTelemetry Collector
-2) SQLite
-3) Console
+Currently, only the meta-reference provider is implemented. It can be configured to send events to multiple sink types:
+1) OpenTelemetry Collector (traces and metrics)
+2) SQLite (traces only)
+3) Console (all events)
 
 #### Configuration
 
-Here's an example that sends telemetry signals to all three sink types. Your configuration might use only one.
+Here's an example that sends telemetry signals to all sink types. Your configuration might use only one or a subset.
+
 ```yaml
   telemetry:
   - provider_id: meta-reference
     provider_type: inline::meta-reference
     config:
+      service_name: "llama-stack-service"
       sinks: ['console', 'sqlite', 'otel_trace', 'otel_metric']
-      otel_trace_endpoint: "http://localhost:4318/v1/traces"
-      otel_metric_endpoint: "http://localhost:4318/v1/metrics"
+      otel_exporter_otlp_endpoint: "http://localhost:4318"
       sqlite_db_path: "/path/to/telemetry.db"
 ```
 
+**Environment Variables:**
+- `OTEL_EXPORTER_OTLP_ENDPOINT`: OpenTelemetry Collector endpoint (default: `http://localhost:4318`)
+- `OTEL_SERVICE_NAME`: Service name for telemetry (default: empty string)
+- `TELEMETRY_SINKS`: Comma-separated list of sinks (default: `console,sqlite`)
+
 ### Jaeger to visualize traces
 
-The `otel` sink works with any service compatible with the OpenTelemetry collector, traces and metrics has two separate endpoints.
-Let's use Jaeger to visualize this data.
+The `otel_trace` sink works with any service compatible with the OpenTelemetry collector. Traces and metrics use separate endpoints but can share the same collector.
 
 Start a Jaeger instance with the OTLP HTTP endpoint at 4318 and the Jaeger UI at 16686 using the following command:
 
@@ -68,4 +137,7 @@ Once the Jaeger instance is running, you can visualize traces by navigating to h
 
 ### Querying Traces Stored in SQLite
 
-The `sqlite` sink allows you to query traces without an external system. Here are some example queries. Refer to the notebook at [Llama Stack Building AI Applications](https://github.com/meta-llama/llama-stack/blob/main/docs/getting_started.ipynb) for more examples on how to query traces and spaces.
+The `sqlite` sink allows you to query traces without an external system. Here are some example
+queries. Refer to the notebook at [Llama Stack Building AI
+Applications](https://github.com/meta-llama/llama-stack/blob/main/docs/getting_started.ipynb) for
+more examples on how to query traces and spans.

docs/source/providers/telemetry/inline_meta-reference.md

@@ -8,10 +8,9 @@ Meta's reference implementation of telemetry and observability using OpenTelemet
 
 | Field | Type | Required | Default | Description |
 |-------|------|----------|---------|-------------|
-| `otel_trace_endpoint` | `str \| None` | No |  | The OpenTelemetry collector endpoint URL for traces |
-| `otel_metric_endpoint` | `str \| None` | No |  | The OpenTelemetry collector endpoint URL for metrics |
+| `otel_exporter_otlp_endpoint` | `str \| None` | No |  | The OpenTelemetry collector endpoint URL (base URL for traces, metrics, and logs). If not set, the SDK will use OTEL_EXPORTER_OTLP_ENDPOINT environment variable. |
 | `service_name` | `<class 'str'>` | No | ​ | The service name to use for telemetry |
-| `sinks` | `list[inline.telemetry.meta_reference.config.TelemetrySink` | No | [<TelemetrySink.CONSOLE: 'console'>, <TelemetrySink.SQLITE: 'sqlite'>] | List of telemetry sinks to enable (possible values: otel, sqlite, console) |
+| `sinks` | `list[inline.telemetry.meta_reference.config.TelemetrySink` | No | [<TelemetrySink.CONSOLE: 'console'>, <TelemetrySink.SQLITE: 'sqlite'>] | List of telemetry sinks to enable (possible values: otel_trace, otel_metric, sqlite, console) |
 | `sqlite_db_path` | `<class 'str'>` | No | ~/.llama/runtime/trace_store.db | The path to the SQLite database to use for storing traces |
 
 ## Sample Configuration
@@ -20,6 +19,7 @@ Meta's reference implementation of telemetry and observability using OpenTelemet
 service_name: "${env.OTEL_SERVICE_NAME:=\u200B}"
 sinks: ${env.TELEMETRY_SINKS:=console,sqlite}
 sqlite_db_path: ${env.SQLITE_STORE_DIR:=~/.llama/dummy}/trace_store.db
+otel_exporter_otlp_endpoint: ${env.OTEL_EXPORTER_OTLP_ENDPOINT:=}
 
 ```
 

llama_stack/providers/inline/telemetry/meta_reference/config.py

@@ -20,13 +20,9 @@ class TelemetrySink(StrEnum):
 
 
 class TelemetryConfig(BaseModel):
-    otel_trace_endpoint: str | None = Field(
+    otel_exporter_otlp_endpoint: str | None = Field(
         default=None,
-        description="The OpenTelemetry collector endpoint URL for traces",
-    )
-    otel_metric_endpoint: str | None = Field(
-        default=None,
-        description="The OpenTelemetry collector endpoint URL for metrics",
+        description="The OpenTelemetry collector endpoint URL (base URL for traces, metrics, and logs). If not set, the SDK will use OTEL_EXPORTER_OTLP_ENDPOINT environment variable.",
     )
     service_name: str = Field(
         # service name is always the same, use zero-width space to avoid clutter
@@ -35,7 +31,7 @@ class TelemetryConfig(BaseModel):
     )
     sinks: list[TelemetrySink] = Field(
         default=[TelemetrySink.CONSOLE, TelemetrySink.SQLITE],
-        description="List of telemetry sinks to enable (possible values: otel, sqlite, console)",
+        description="List of telemetry sinks to enable (possible values: otel_trace, otel_metric, sqlite, console)",
     )
     sqlite_db_path: str = Field(
         default_factory=lambda: (RUNTIME_BASE_DIR / "trace_store.db").as_posix(),
@@ -55,4 +51,5 @@ class TelemetryConfig(BaseModel):
             "service_name": "${env.OTEL_SERVICE_NAME:=\u200b}",
             "sinks": "${env.TELEMETRY_SINKS:=console,sqlite}",
             "sqlite_db_path": "${env.SQLITE_STORE_DIR:=" + __distro_dir__ + "}/" + db_name,
+            "otel_exporter_otlp_endpoint": "${env.OTEL_EXPORTER_OTLP_ENDPOINT:=}",
         }

llama_stack/providers/inline/telemetry/meta_reference/telemetry.py

@@ -86,24 +86,27 @@ class TelemetryAdapter(TelemetryDatasetMixin, Telemetry):
             provider = TracerProvider(resource=resource)
             trace.set_tracer_provider(provider)
             _TRACER_PROVIDER = provider
-            if TelemetrySink.OTEL_TRACE in self.config.sinks:
-                if self.config.otel_trace_endpoint is None:
-                    raise ValueError("otel_trace_endpoint is required when OTEL_TRACE is enabled")
-                span_exporter = OTLPSpanExporter(
-                    endpoint=self.config.otel_trace_endpoint,
-                )
-                span_processor = BatchSpanProcessor(span_exporter)
-                trace.get_tracer_provider().add_span_processor(span_processor)
-            if TelemetrySink.OTEL_METRIC in self.config.sinks:
-                if self.config.otel_metric_endpoint is None:
-                    raise ValueError("otel_metric_endpoint is required when OTEL_METRIC is enabled")
-                metric_reader = PeriodicExportingMetricReader(
-                    OTLPMetricExporter(
-                        endpoint=self.config.otel_metric_endpoint,
+
+            # Use single OTLP endpoint for all telemetry signals
+            if TelemetrySink.OTEL_TRACE in self.config.sinks or TelemetrySink.OTEL_METRIC in self.config.sinks:
+                if self.config.otel_exporter_otlp_endpoint is None:
+                    raise ValueError(
+                        "otel_exporter_otlp_endpoint is required when OTEL_TRACE or OTEL_METRIC is enabled"
                     )
-                )
-                metric_provider = MeterProvider(resource=resource, metric_readers=[metric_reader])
-                metrics.set_meter_provider(metric_provider)
+
+                # Let OpenTelemetry SDK handle endpoint construction automatically
+                # The SDK will read OTEL_EXPORTER_OTLP_ENDPOINT and construct appropriate URLs
+                # https://opentelemetry.io/docs/languages/sdk-configuration/otlp-exporter
+                if TelemetrySink.OTEL_TRACE in self.config.sinks:
+                    span_exporter = OTLPSpanExporter()
+                    span_processor = BatchSpanProcessor(span_exporter)
+                    trace.get_tracer_provider().add_span_processor(span_processor)
+
+                if TelemetrySink.OTEL_METRIC in self.config.sinks:
+                    metric_reader = PeriodicExportingMetricReader(OTLPMetricExporter())
+                    metric_provider = MeterProvider(resource=resource, metric_readers=[metric_reader])
+                    metrics.set_meter_provider(metric_provider)
+
             if TelemetrySink.SQLITE in self.config.sinks:
                 trace.get_tracer_provider().add_span_processor(SQLiteSpanProcessor(self.config.sqlite_db_path))
             if TelemetrySink.CONSOLE in self.config.sinks:

llama_stack/templates/meta-reference-gpu/run-with-safety.yaml

@@ -64,6 +64,7 @@ providers:
       service_name: "${env.OTEL_SERVICE_NAME:=\u200B}"
       sinks: ${env.TELEMETRY_SINKS:=console,sqlite}
       sqlite_db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/meta-reference-gpu}/trace_store.db
+      otel_exporter_otlp_endpoint: ${env.OTEL_EXPORTER_OTLP_ENDPOINT:=}
   eval:
   - provider_id: meta-reference
     provider_type: inline::meta-reference

llama_stack/templates/meta-reference-gpu/run.yaml

@@ -54,6 +54,7 @@ providers:
       service_name: "${env.OTEL_SERVICE_NAME:=\u200B}"
       sinks: ${env.TELEMETRY_SINKS:=console,sqlite}
       sqlite_db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/meta-reference-gpu}/trace_store.db
+      otel_exporter_otlp_endpoint: ${env.OTEL_EXPORTER_OTLP_ENDPOINT:=}
   eval:
   - provider_id: meta-reference
     provider_type: inline::meta-reference

llama_stack/templates/open-benchmark/run.yaml

@@ -73,6 +73,7 @@ providers:
       service_name: "${env.OTEL_SERVICE_NAME:=\u200B}"
       sinks: ${env.TELEMETRY_SINKS:=console,sqlite}
       sqlite_db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/open-benchmark}/trace_store.db
+      otel_exporter_otlp_endpoint: ${env.OTEL_EXPORTER_OTLP_ENDPOINT:=}
   eval:
   - provider_id: meta-reference
     provider_type: inline::meta-reference

llama_stack/templates/starter/run.yaml

@@ -193,6 +193,7 @@ providers:
       service_name: "${env.OTEL_SERVICE_NAME:=\u200B}"
       sinks: ${env.TELEMETRY_SINKS:=console,sqlite}
       sqlite_db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter}/trace_store.db
+      otel_exporter_otlp_endpoint: ${env.OTEL_EXPORTER_OTLP_ENDPOINT:=}
   post_training:
   - provider_id: huggingface
     provider_type: inline::huggingface

llama_stack/templates/vllm-gpu/run.yaml

@@ -53,6 +53,7 @@ providers:
       service_name: "${env.OTEL_SERVICE_NAME:=\u200B}"
       sinks: ${env.TELEMETRY_SINKS:=console,sqlite}
       sqlite_db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/vllm-gpu}/trace_store.db
+      otel_exporter_otlp_endpoint: ${env.OTEL_EXPORTER_OTLP_ENDPOINT:=}
   eval:
   - provider_id: meta-reference
     provider_type: inline::meta-reference

llama_stack/templates/watsonx/run.yaml

@@ -50,6 +50,7 @@ providers:
       service_name: "${env.OTEL_SERVICE_NAME:=\u200B}"
       sinks: ${env.TELEMETRY_SINKS:=console,sqlite}
       sqlite_db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/watsonx}/trace_store.db
+      otel_exporter_otlp_endpoint: ${env.OTEL_EXPORTER_OTLP_ENDPOINT:=}
   eval:
   - provider_id: meta-reference
     provider_type: inline::meta-reference

scripts/setup_telemetry.sh

@@ -0,0 +1,121 @@
+#!/usr/bin/env bash
+
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+# Telemetry Setup Script for Llama Stack
+# This script sets up Jaeger, OpenTelemetry Collector, Prometheus, and Grafana using Podman
+# For whoever is interested in testing the telemetry stack, you can run this script to set up the stack.
+#    export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318
+#    export TELEMETRY_SINKS=otel_trace,otel_metric
+#    export OTEL_SERVICE_NAME=my-llama-app
+# Then run the distro server
+
+set -Eeuo pipefail
+
+CONTAINER_RUNTIME=${CONTAINER_RUNTIME:-docker}
+
+echo "🚀 Setting up telemetry stack for Llama Stack using Podman..."
+
+if ! command -v "$CONTAINER_RUNTIME" &> /dev/null; then
+  echo "🚨 $CONTAINER_RUNTIME could not be found"
+  echo "Docker or Podman is required. Install Docker: https://docs.docker.com/get-docker/ or Podman: https://podman.io/getting-started/installation"
+  exit 1
+fi
+
+# Create a network for the services
+echo "📡 Creating $CONTAINER_RUNTIME network..."
+$CONTAINER_RUNTIME network create llama-telemetry 2>/dev/null || echo "Network already exists"
+
+# Stop and remove existing containers
+echo "🧹 Cleaning up existing containers..."
+$CONTAINER_RUNTIME stop jaeger otel-collector prometheus grafana 2>/dev/null || true
+$CONTAINER_RUNTIME rm jaeger otel-collector prometheus grafana 2>/dev/null || true
+
+# Start Jaeger
+echo "🔍 Starting Jaeger..."
+$CONTAINER_RUNTIME run -d --name jaeger \
+  --network llama-telemetry \
+  -e COLLECTOR_ZIPKIN_HOST_PORT=:9411 \
+  -p 16686:16686 \
+  -p 14250:14250 \
+  -p 9411:9411 \
+  docker.io/jaegertracing/all-in-one:latest
+
+# Start OpenTelemetry Collector
+echo "📊 Starting OpenTelemetry Collector..."
+$CONTAINER_RUNTIME run -d --name otel-collector \
+  --network llama-telemetry \
+  -p 4318:4318 \
+  -p 4317:4317 \
+  -p 9464:9464 \
+  -p 13133:13133 \
+  -v $(pwd)/otel-collector-config.yaml:/etc/otel-collector-config.yaml:Z \
+  docker.io/otel/opentelemetry-collector-contrib:latest \
+  --config /etc/otel-collector-config.yaml
+
+# Start Prometheus
+echo "📈 Starting Prometheus..."
+$CONTAINER_RUNTIME run -d --name prometheus \
+  --network llama-telemetry \
+  -p 9090:9090 \
+  -v $(pwd)/prometheus.yml:/etc/prometheus/prometheus.yml:Z \
+  docker.io/prom/prometheus:latest \
+  --config.file=/etc/prometheus/prometheus.yml \
+  --storage.tsdb.path=/prometheus \
+  --web.console.libraries=/etc/prometheus/console_libraries \
+  --web.console.templates=/etc/prometheus/consoles \
+  --storage.tsdb.retention.time=200h \
+  --web.enable-lifecycle
+
+# Start Grafana
+echo "📊 Starting Grafana..."
+$CONTAINER_RUNTIME run -d --name grafana \
+  --network llama-telemetry \
+  -p 3000:3000 \
+  -e GF_SECURITY_ADMIN_PASSWORD=admin \
+  -e GF_USERS_ALLOW_SIGN_UP=false \
+  docker.io/grafana/grafana:latest
+
+# Wait for services to start
+echo "⏳ Waiting for services to start..."
+sleep 10
+
+# Check if services are running
+echo "🔍 Checking service status..."
+$CONTAINER_RUNTIME ps --filter "name=jaeger|otel-collector|prometheus|grafana" --format "table {{.Names}}\t{{.Status}}\t{{.Ports}}"
+
+echo ""
+echo "✅ Telemetry stack is ready!"
+echo ""
+echo "🌐 Service URLs:"
+echo "   Jaeger UI:        http://localhost:16686"
+echo "   Prometheus:       http://localhost:9090"
+echo "   Grafana:          http://localhost:3000 (admin/admin)"
+echo "   OTEL Collector:   http://localhost:4318 (OTLP endpoint)"
+echo ""
+echo "🔧 Environment variables for Llama Stack:"
+echo "   export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318"
+echo "   export TELEMETRY_SINKS=otel_trace,otel_metric"
+echo "   export OTEL_SERVICE_NAME=my-llama-app"
+echo ""
+echo "📊 Next steps:"
+echo "   1. Set the environment variables above"
+echo "   2. Start your Llama Stack application"
+echo "   3. Make some inference calls to generate metrics"
+echo "   4. Check Jaeger for traces: http://localhost:16686"
+echo "   5. Check Prometheus for metrics: http://localhost:9090"
+echo "   6. Set up Grafana dashboards: http://localhost:3000"
+echo ""
+echo "🔍 To test the setup, run:"
+echo "   curl -X POST http://localhost:5000/v1/inference/chat/completions \\"
+echo "     -H 'Content-Type: application/json' \\"
+echo "     -d '{\"model_id\": \"your-model\", \"messages\": [{\"role\": \"user\", \"content\": \"Hello\"}]}'"
+echo ""
+echo "🧹 To clean up when done:"
+echo "   $CONTAINER_RUNTIME stop jaeger otel-collector prometheus grafana"
+echo "   $CONTAINER_RUNTIME rm jaeger otel-collector prometheus grafana"
+echo "   $CONTAINER_RUNTIME network rm llama-telemetry"