Rolling swap can start the new replica before a migration finishes #94

New issue

Closed

opened 2026-10-10 05:28:27 +00:00 by ahmad · 0 comments

ahmad commented

2026-10-10 05:28:27 +00:00

Owner

Context

With rolling-services, deploy-vps recreates the other services with --no-deps and then rolls the listed ones, also with --no-deps. up -d returns as soon as a one-shot service such as a migration has started. So the new replica can start before the migration finishes, and even after it has failed. Without rolling, compose's depends_on: condition: service_completed_successfully prevents that.

Found while adopting the swap in suhba/rihla#417, and tested on the VPS:

rolling without --no-deps waits correctly, but re-runs a migration that has already exited (twice per deploy);
up --wait reports the running one-shot as healthy and returns before it ends.

Scope

In: in rolling mode, before any swap, wait for each recreated container whose restart policy is no to exit, and fail the deploy (old replicas untouched) unless it exits 0. Document it.
Out: the non-rolling path, which keeps today's behaviour.

Acceptance criteria

Given a rolling deploy whose migration takes a while, when it runs, then the new replica starts only after the migration has exited 0, and the migration runs once.
Given a migration that fails, when deploying, then no new replica starts, the old one keeps serving, and the deploy fails with the migration's logs.
Given a consumer that does not set rolling-services, when it deploys, then nothing changes.

### Context With `rolling-services`, deploy-vps recreates the other services with `--no-deps` and then rolls the listed ones, also with `--no-deps`. `up -d` returns as soon as a one-shot service such as a migration has started. So the new replica can start before the migration finishes, and even after it has failed. Without rolling, compose's `depends_on: condition: service_completed_successfully` prevents that. Found while adopting the swap in suhba/rihla#417, and tested on the VPS: - rolling without `--no-deps` waits correctly, but re-runs a migration that has already exited (twice per deploy); - `up --wait` reports the running one-shot as healthy and returns before it ends. ### Scope - In: in rolling mode, before any swap, wait for each recreated container whose restart policy is `no` to exit, and fail the deploy (old replicas untouched) unless it exits 0. Document it. - Out: the non-rolling path, which keeps today's behaviour. ### Acceptance criteria - [ ] Given a rolling deploy whose migration takes a while, when it runs, then the new replica starts only after the migration has exited 0, and the migration runs once. - [ ] Given a migration that fails, when deploying, then no new replica starts, the old one keeps serving, and the deploy fails with the migration's logs. - [ ] Given a consumer that does not set `rolling-services`, when it deploys, then nothing changes.

ahmad added the

labels

2026-10-10 05:28:27 +00:00

ahmad self-assigned this

2026-10-10 05:28:27 +00:00

ahmad referenced this issue from a pull request that will close it,

2026-10-10 05:31:50 +00:00

fix: Wait for one-shots before a rolling swap #95