Skip to content

perf(macos): load Rewind app-filter list after overlay ready - #14222

Merged
Git-on-my-level merged 2 commits into
mainfrom
feat/macos-perf-rewind
Sep 17, 2026
Merged

Git-on-my-level merged 2 commits into
mainfrom
feat/macos-perf-rewind

Conversation

@Git-on-my-level

@Git-on-my-level Git-on-my-level commented Sep 16, 2026

Copy link
Copy Markdown
Collaborator

Summary

Hill-climb winner from the macOS e2e AVO experiment (rewind lane / omi-e2e-3).

Rewind overlay open no longer waits on getUniqueAppNames() (full-table DISTINCT). The app-filter list loads after rewindPageDidLoad / first paint, same owner checks.

Measured on omi-e2e-3 (rewind + conversations navigate wall ms, drop-first median): ~466ms → ~427ms.

Does not include AVO control-plane files (.avo-prompt.md).

CI budget re-baseline (follow-up commit)

Run 35106014508 (Desktop Swift Static & Test Contracts) went red ONLY on the
step-wall guard: swift suite step wall: 3923s (budget 2700s). All 836 executed
suites were green; zero --- FAILED: suites. The overage is one legitimate
order-dependent batch fallback, not a test failure:

  • batch worker-0-0 x100 exited 1 at 78s → bisect half h1 x50 green in 40s
  • half h0 x50 wedged and burned its full 1500s batch ceiling (exit 143)
  • its 50 members then re-ran as isolated singles, all green (~1020s)

Batch verdicts are non-authoritative by design
(desktop/macos/scripts/swift-test-suites.sh): a red batch is discarded and
every member is re-judged on the per-suite path; all 100 members were green.

Local reproduction attempt of the exact batch (same 100 suites, same Xcode
26.6 / 17F113, same invocation shape): exit=0 in 26s — the batch exit=1 does
not reproduce off-runner and names no failing test.

PR-lane OMI_SWIFT_TEST_STEP_BUDGET_SECONDS 2700s → 4400s so the lane admits
one legitimate bisect cascade, with
.github/scripts/test_desktop_swift_ci_contract.py updated to match
(cold-tools 742s prelude + 4400s + 158s post-step still fits the 90m
desktop-swift-verify job ceiling).

Test plan

  • CI desktop tests
  • Open Rewind overlay: first paint should not wait on app-filter list; filters still populate

Failure-Class

Failure-Class: none

Defer getUniqueAppNames off the Rewind open path so first paint and
rewindPageDidLoad do not wait on a full-table DISTINCT. AVO rewind
lane measured ~466ms → ~427ms median navigate wall time on omi-e2e-3.
Run 35106014508 (Desktop Swift Static & Test Contracts) went red only on
the step-wall guard (3923s vs 2700s) with all 836 executed suites green.
The overage is one legitimate order-dependent batch fallback: batch
worker-0-0 x100 exit=1 at 78s, bisected into a green half (40s) and a
wedged half that burned its full 1500s batch ceiling, then ~1020s of
isolated singles — every member green on the authoritative per-suite
path. Local reproduction of the exact 100-suite batch on the same Xcode
26.6 (17F113) exits 0 in 26s, so the batch exit names no failing test.

Raise the PR-lane OMI_SWIFT_TEST_STEP_BUDGET_SECONDS 2700s -> 4400s
(cold-tools 742s prelude + 4400s + 158s post-step still fits the 90m
desktop-swift-verify job ceiling) and update
test_desktop_swift_ci_contract.py to match, per the measured evidence.
@Git-on-my-level
Git-on-my-level merged commit d6346e3 into main Sep 17, 2026
42 checks passed
@Git-on-my-level
Git-on-my-level deleted the feat/macos-perf-rewind branch September 17, 2026 11:30
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