metrics

Google ADK model usage metrics

This sample records successful Google ADK model calls, model errors, and token usage from LLMResponse.UsageMetadata. The ADK callbacks emit through Temporal's Workflow metrics handler, which suppresses emissions during replay, and tag each counter with fixed agent and model names.

Counter (as scraped)	Source
google_adk_model_calls_total	One per successful model turn
google_adk_model_errors_total	One per failed model turn
google_adk_input_tokens_total	PromptTokenCount
google_adk_output_tokens_total	CandidatesTokenCount
google_adk_reasoning_tokens_total	ThoughtsTokenCount
google_adk_cached_input_tokens_total	CachedContentTokenCount
google_adk_total_tokens_total	TotalTokenCount

Prometheus adds the _total suffix and sanitizes tag values. The SDK also adds a workflow_type tag, so a scraped line looks like google_adk_model_calls_total{agent="metrics_assistant",model="gemini_2_5_flash",workflow_type="AgentWorkflow"} 1.

Replay suppression makes these metrics at-least-once rather than exactly-once. A Workflow Task retry re-executes its live segment and can record again, so use these as usage signals rather than billing records.

Start a local Temporal Service:

temporal server start-dev

Set a Gemini API key and run the worker:

export GEMINI_API_KEY=...
go run googleadk/metrics/worker/main.go

In another terminal, start the Workflow:

go run googleadk/metrics/starter/main.go

Inspect the Prometheus endpoint:

curl -s http://127.0.0.1:9090/metrics | grep google_adk_