Skip to main content

Production & Deployment

🎯 Purpose: Deploy ADK agents to production with proper observability, scaling, and service management.

📚 Source of Truth: google/adk-python/src/google/adk/cli/ (ADK 1.15) + Official deployment guides

🏭 Deployment Environments​

Local Development​

# Quick development testing
adk web agent_name

# Run with custom config
adk run agent_name --config config.yaml

Cloud Run (Serverless)​

# Deploy to Cloud Run
adk deploy cloud_run agent_name

# Auto-scaling, pay-per-use
# Services: Cloud SQL, GCS, Vertex AI

Vertex AI Agent Engine (Managed)​

# Enterprise deployment
adk deploy agent_engine agent_name

# Fully managed by Google
# High availability, monitoring

GKE (Kubernetes)​

# Custom infrastructure
adk deploy gke agent_name

# Full control, custom scaling

[CALLB] Observability & Monitoring​

Events (What Happened)​

# Enable event logging
runner = Runner(
    event_service=LoggingEventService(level="DEBUG")
)

# Events captured:
# - AGENT_START/COMPLETE
# - TOOL_CALL_START/RESULT
# - LLM_REQUEST/RESPONSE
# - STATE_CHANGE

Tracing (Why It Happened)​

# Detailed execution traces
runner = Runner(
    trace_service=CloudTraceService(project="my-project")
)

# View in Cloud Trace console
# Performance bottlenecks
# Error root causes

Callbacks (Custom Monitoring)​

def monitor_agent(context, result):
    # Custom metrics
    log_performance(result.execution_time)
    alert_on_errors(result.errors)

agent = Agent(
    name="monitored_agent",
    callbacks=[monitor_agent]
)

Evaluation (Quality Metrics)​

# Automated testing
adk eval agent_name --test-set my_tests.evalset.json

# Metrics:
# - tool_trajectory_avg_score (0-1)
# - response_match_score (0-1)
# - Custom LLM-as-judge metrics

💾 Service Configuration​

Development (InMemory)​

runner = Runner()  # All services default to InMemory

Production (Persistent)​

runner = Runner(
    session_service=PostgresSessionService(uri="..."),
    artifact_service=GcsArtifactService(bucket="..."),
    memory_service=VertexAiMemoryBankService(project="...")
)

🔒 Security & Best Practices​

  • Environment Variables: Never commit secrets
  • Service Accounts: Least privilege access
  • Input Validation: Sanitize all inputs
  • Rate Limiting: Protect against abuse
  • Error Handling: Graceful failure modes

📊 Performance Optimization​

  • Model Selection: Right model for cost/performance
  • Caching: Reuse expensive computations
  • Parallel Execution: Independent tasks simultaneously
  • Batch Processing: Group similar requests

🎯 Key Takeaways​

  1. Multiple deployment options: Local, Cloud Run, Vertex AI, GKE
  2. Observability layers: Events, traces, callbacks, evaluation
  3. Service configuration: InMemory for dev, persistent for prod
  4. Security first: Environment variables, validation, rate limiting
  5. Performance: Optimize models, caching, parallel execution

🔗 Next: Explore Advanced Patterns for cutting-edge capabilities.