Support singleton deploy role for services like Celery Beat

[wadewilliams]

Problem

Services like Celery Beat are singletons — only one instance should run at a time. The current deploy roles don't fit well:

• app — rolling deploy scales to 2 temporarily, causing duplicate scheduling
• accessory — never restarted, so it runs stale code after deploy

Current workaround

Using deploy.role=app with deploy.drain=1 to minimize the overlap window. Tasks are assumed idempotent, so brief duplicate scheduling is tolerable.

Proposal

A third deploy role (e.g. singleton or sequential) that does stop-before-start instead of rolling:

1. Send SIGTERM to the old container, wait for graceful shutdown
2. Start the new container
3. Health check the new container

This would be useful for any service where exactly-one-at-a-time semantics matter: beat schedulers, cron runners, lock holders, single-consumer queues.

Alternatives considered

• celery-redbeat: Distributed lock scheduler that makes beat safe for rolling deploys. Works but adds a dependency to solve a deploy-tool gap.
• deploy.drain=1: Minimizes overlap but doesn't eliminate it.