Skip to content

Deployment 🚀

Getting a prototype running locally is great; putting it in production requires reliability, observability, and scalability. Flock ships with first-class Temporal integration which gives you all three out of the box.

1. Local vs. Temporal

Aspect Local Mode (default) Temporal Mode
Fault Tolerance Exceptions bubble to caller Automatic retries, timeouts, watchdogs
State Persistence Memory-only Durable (DB, visibility store)
Concurrency Single Python process Horizontal scaling across workers
Visibility Stdout + logs Temporal Web UI, Prometheus, Grafana

Switching modes is as easy as:

flock = Flock(enable_temporal=True, temporal_config=TemporalWorkflowConfig(...))

2. Production Checklist

  1. Secrets Management – Store LLM API keys in Vault, AWS Parameter Store, etc.
  2. Observability – Export OpenTelemetry traces to Jaeger or Tempo.
  3. Autoscaling – Use temporal worker --concurrency or K8s HPA.
  4. Caching – Enable evaluator cache (Redis, Memcached, or litellm built-in).
  5. Security – Restrict tool access; disable cloudpickle deserialization if loading untrusted payloads.

3. Reference Architectures

Docker Compose (PoC)

  • api (FastAPI) + flock code
  • Temporal server (Cadence) or temporal-io docker image
  • Prometheus + Grafana for metrics

Kubernetes (Prod)

  • Flock API image
  • Temporal Cluster (6-service helm chart)
  • Workers scaled by queue
  • OpenTelemetry Collector → Tempo → Grafana

4. Next Steps

  • Learn the details in Temporal Deployment.
  • Check the scripts/ folder for helper shell scripts.
  • Browse tests/ for smoke tests that validate a cluster.