Jobs & hybrid cloud

Jobs (run-to-completion)

Not everything is a long-running server. A kind: job runs once and its dependents wait for it to exit 0 — completion-gated, not health-gated. This is the shape for database migrations, seed data, and ETL/ML pipeline steps.

services:
  - { name: db, docker: { image: postgres:16 }, health: { tcp: "localhost:5432" } }
  - name: migrate
    kind: job
    command: "alembic upgrade head"
    dependsOn: [db]
  - name: api
    command: "uvicorn app:api"
    dependsOn: [migrate]      # starts only after migrate exits 0

A failed job fails the run instead of starting its dependents. Chain jobs with dependsOn to express a pipeline.

Local + cloud (hybrid)

Real products often need things that don't run locally — a managed database, a SaaS API, a service that only lives in a dev cluster.

• External dependencies (external: true): not started, only health-probed for reachability. If the cloud DB is down you find out before everything else boots.
• Tunnels are just services. A kubectl port-forward or ssh -L is a long-running command — model it as a service with restart: always and a health.tcp on the local port.
• Lifecycle hooks run global one-shot commands around the environment:

hooks:
  preUp:   ["gcloud auth login", "fetch-secrets > .env.cloud"]
  postDown: ["echo done"]