[etareduction]

CI on main intermittently ends with N tests, 0 failures, 2–5 invalid: one module's setup_all fails with

** (RuntimeError) could not lookup Ecto repo AshOban.Test.Repo because it was not started or it does not exist
  ...
  (oban 2.24.1) lib/oban/migration.ex:212: Oban.Migration.verify_migrated!/1
  (oban 2.24.1) lib/oban.ex:599: Oban.start_link/1

(e.g. runs 39450958371, 38788849819, 38503023987, 33346691045, 33293567334). The module varies.

Seven test modules call AshOban.Test.Repo.start_link() in setup_all and ignore the result. The repo gets linked to the setup_all process of whichever module started it first, and ExUnit only stops that process's linked children asynchronously after the module finishes. The next module's start_link() can return {:error, {:already_started, _}} just before the old repo goes down, and then Oban.start_link/1 finds no repo.

This starts the repo once in test_helper.exs and drops the per-module calls.

Reproducing

--repeat-until-failure runs the suite repeatedly in one VM. VerifyUseTenantFromRecordTest defines modules inside its tests and fails on any second iteration, so it has to be excluded:

mix test --seed 564897 --repeat-until-failure 500 --exclude module:AshOban.Verifiers.VerifyUseTenantFromRecordTest

On main, 11 out of 11 trials (seeds 564897 and 1–10) produced invalid modules, after 12–289 iterations (median 31). Across 992 iterations that is about 1 in 90, so 500 repeats miss it about 0.4% of the time. With this change, seeds 564897 and 1–3 ran all 501 iterations each (2004 in total) with no failures or invalid modules.

[zachdaniel]

🚀 Thank you for your contribution! 🚀