Skip to content

land: watcher-retry-test - #757

Merged
ualtinok merged 2 commits into
mainfrom
land/watcher-retry-test
Oct 22, 2026
Merged

ualtinok merged 2 commits into
mainfrom
land/watcher-retry-test

Conversation

@aft-alfonso

@aft-alfonso aft-alfonso Bot commented Oct 22, 2026 •

Copy link
Copy Markdown

Accepted agent delivery; branch-keyed search descriptor included.


Summary by cubic

Fixes exclusion_planner_timed_out_roots_retry_same_generation, which timed out on loaded Windows CI runners: the 2 s test budget covers both waiting for a pool permit and the filesystem plan, so retries queued behind the pool's two permits ran out of time. The test now waits for the released workers to return their permits, then retries each root sequentially with the plan computed before the attempt. Retry bookkeeping assertions are unchanged.

Adds a search-quality benchmark descriptor for watcher-retry-test.

Written for commit 97eaf8f. Summary will update on new commits.

View guided diff Turn on auto-fix

RetriggerView in GreptileConfidence Score: 5/5

This PR appears safe to merge.

Summary

This PR stabilizes a watcher retry test without changing production behavior.

  • The retry test waits for released planner workers to finish.
  • Each retry gets a ready plan before its time limit starts.

Reviews (1) · Last reviewed commit: "release notes and search-quality descrip..." · Reviewed by Greptile

ualtinok and others added 2 commits October 22, 2026 20:32
…dget

exclusion_planner_timed_out_roots_retry_same_generation failed on Windows
CI with "capacity is free: IoDeadline". After the test released its two held
workers, eight retries ran at once through the pool's two permits. Each
retry had a 2 s budget, and that budget covers both the wait for a permit
and the filesystem plan. On a loaded Windows runner one plan of these small
roots took 460 ms to 1.23 s (about 1 ms when the runner is idle), so the
retries at the back of the queue ran out of time. The retries' premise,
that capacity is free, was false.

This is a test bug. The planner is not compiled into Windows production
builds, and on Linux and macOS an unavailable plan is retried with backoff
by design.

The test now waits until the released workers return their permits, then
retries each root in turn with a plan computed before the attempt. The
attempt still takes a permit from the shared pool, and the retry
bookkeeping assertions are unchanged.

Load loop on the native Windows runner, 8 parallel copies beside the full
lib suite: 8 failures in 2400 runs before, 0 in 2400 after. Mutation: making
record_outcome consume the generation on failure reddens this test only.

Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
(cherry picked from commit f123bd09cceb8f3bc4253497e31a417b7cd5ea80)
Co-authored-by: alfonso-aft <289616620+alfonso-aft@users.noreply.github.com>
@ualtinok
ualtinok added this pull request to the merge queue Oct 22, 2026
Merged via the queue into main with commit 30aaac9 Oct 22, 2026
83 of 84 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant