README.md

Google ADK replay-safe metrics

This sample exports Google ADK's OpenTelemetry metrics to a local Prometheus endpoint while preventing Workflow replay from recording the same observations again. The default scripted model is deterministic and makes no network model calls, so no API key is needed.

Start a local Temporal development server:

temporal server start-dev

In another terminal, start the worker from the repository root:

uv run python -m google_adk_agents.metrics.run_worker

Then run the Workflow:

uv run python -m google_adk_agents.metrics.run_metrics_workflow

The starter prints Replay-safe metrics are ready. Inspect the metrics exposed by the worker:

curl -s http://127.0.0.1:9464/metrics | grep gen_ai

The output includes gen_ai.invoke_agent, gen_ai.client.operation.duration, and gen_ai.client.token.usage metrics. Prometheus replaces dots with underscores, so an exported line looks like gen_ai_invoke_agent_duration_seconds_count{gen_ai_agent_name="metrics_agent"} 1.0. ReplaySafeMeterProvider drops observations made while replaying, so replay does not multiply the recorded counts.

Recordings are first-execution-only rather than exactly-once. Replay is suppressed, but a Workflow Task retry re-executes live and can record again, so treat these metrics as at-least-once usage signals.

OpenTelemetry's global meter provider can be installed only once per process. run_worker.py installs the replay-safe provider before importing Google ADK or the Workflow. Applications embedding this setup must likewise make it the first and only global meter provider installation in that process.