[earakely-scale]

Why

After the slow tier went to eight runners, the longest integration jobs on main were the fast tier, integration-local, at 15.9 minutes, and the two gateway shards at 12.8 and 11.9 minutes. The slow shards took 5.5–8.8 minutes. Most of that time goes to deploys rather than to the tests themselves.

What

• The trigger tests deploy once. The five tests in test_register_env_triggers_local.py share one deployed items env and its registry, made by a module fixture, suite.
  • A per-test fixture, items_env, returns the env to the state a fresh deploy leaves it in:
    • POST /triggers/clear removes triggers, events, watch roles, executors and config.
    • POST /clock/clear stops the clock.
    • The env's data/reset empties the items.
  • The trajectory is append-only, so the delay-and-hours test reads only the lines added after it starts.
  • The stack setup in local_parallel_deploy_test.py moves into a context manager, local_stack_at, which a module can hold for all its tests. local_stack still gives one test its own.
• Two fast-tier files move to the slow tier. env/envs/server_env_local_test.py and openenv/build_app_test.py deploy envs in nearly every test. They ran first in the fast tier, on a cold image cache, and took about 8.6 of its minutes. They are now int_test_slow, as the marker's definition says: real Docker builds and deploys. In the slow shards, where earlier modules have warmed the image cache, they take 72 and 79 seconds.
• The gateway tests run on three runners, not two: integration-local-gateway (1/3) to (3/3). The required checks keep their names.
• slow_durations.json is refreshed from this branch's own run (40642453175), so the shards balance with the moved files and the new trigger times.

Tests

CI on this branch, compared with main at f6ae9cc:

Job	main	This branch
integration-local (fast tier)	15.9 min	5.6 min
integration-local-gateway shards	11.9, 12.8 min	9.5, 8.7, 9.8 min
integration-local-slow shards	5.5–8.8 min	6.4–8.2 min

The longest integration job goes from 15.9 to 9.8 minutes. unit (7.5 minutes) is unchanged.

The trigger tests on their own, on a 4-vCPU Ubuntu 24.04 VM set up as the slow-tier job sets up its runner:

Run	Time	Result
main, warm image cache	268 s, 50–59 s per test	5 passed
This branch, warm	89 s: 43 s for the one deploy, 5–16 s per test	5 passed
This branch, the five tests in reverse order	89 s	5 passed
This branch, followed by local_parallel_deploy_test.py in the same session	154 s	6 passed

No test containers were left behind after any run. ruff check and tst/unit/test_shards.py pass.

🤖 Generated with Claude Code