Two run tests assumed timings a hosted runner does not give (0.6.8)

[sotashimozono]

DataVault.jl's downstream SweepRunner job failed 3 times out of 15 identical CompatHelper PRs (#60, #62, #72). Each failure was one of these two assertions, and the contents of those PRs were identical to the 12 that passed. The EOFError / Error encountered while load … manifest.jld2 lines in those logs appear in passing runs too, the same number of times, and are not the failure.

test_run_loop_busy.jl:51 — elapsed >= 3.0 (failed at 2.85 s and 2.90 s)

DataVault writes heartbeat= truncated to the second ("yyyy-mm-ddTHH:MM:SS"), so a lock reads up to 1 s older than it is and is reclaimed up to 1 s before stale_after.

Measured locally:

lock written at	age right after writing	age > 3.0 first seen at
.05 s	0.06 s	2.96 s
.95 s	0.94 s	2.07 s, 2.08 s

When elapsed lands between 2 and 3 s depends on where the round boundaries fall, which is why the test failed only sometimes. The bound is now stale_after − 1.

Does it still catch the bug it pins? With the busy wait disabled (if false && result.busy > 0), run_loop! returned in 0.96 / 0.99 / 1.01 s with done == 0 after warm-up, so >= 2.0 fails as it should. r.done == 1 and is_done also catch it. The first cold run took 9.7 s because of compilation and reclaimed the key; the old >= 3.0 misses that case too, so the new bound is no weaker.

With stale_after = 600 in production, reclaiming up to 1 s early does not matter. The .running format is a contract, so DataVault is not changed.

test_run_deadline.jl:58 — started[] >= 1 (was 0)

A deadline 0.3 s ahead passed before run! handed out its first key: a run-start observation takes most of a second. This is the same cause as e207f53. RunOpts is immutable and deadline is absolute, so the fix cannot trigger on the first key the way that commit did. Instead the deadline is 5 s ahead and the first key runs until it has passed. The test still checks the same thing (at least one key started, not every key, returned after the deadline, stopped_by === :deadline) and no longer depends on how fast run! starts. It takes about 5 s longer.

Checked

• Both files pass locally in a throwaway env (test_run_loop_busy.jl 23 s, test_run_deadline.jl 19 s). JuliaFormatter 2 reports them already formatted.

🤖 Generated with Claude Code

[github-actions]

📚 Docs preview: https://qatlashub.github.io/SweepRunner.jl/previews/PR60/

(updates on each push to this PR)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!