isaacbmiller · 2026-08-04T18:13:03Z

Summary

Ports the completed-result tracking portion of dbreunig#4 as an independent upstream change.

use a private sentinel to distinguish unfinished slots from legitimate None results
count None-returning tasks as completed in sequential and threaded progress
prevent completed None slots from being accepted again after straggler resubmission
preserve the public result contract where failed or unfinished slots are returned as None

The original PR’s SIGINT-handler change is intentionally excluded because guarding with callable() does not define correct behavior for SIG_DFL versus SIG_IGN.

Original implementation authored by @dbreunig, with a progress-accounting regression assertion added during upstream review.

Test plan

pytest tests/utils/test_parallelizer.py -q (14 passed)
git diff --check
upstream CI

Lint note

A local whole-file Ruff run found two pre-existing RUF043 findings in test_parallelizer.py lines 62 and 147; this PR does not modify those assertions.

Co-authored-by: Isaac Miller <isaacbmiller@users.noreply.github.com> Amp-Thread-ID: https://ampcode.com/threads/T-019fcd6b-0068-752e-b9ce-2fd40b7f061b Co-authored-by: Isaac Miller <isaac@cmpnd.ai>

greptile-apps · 2026-08-04T18:19:41Z

Greptile Summary

This update makes ParallelExecutor distinguish unfinished result slots from tasks that legitimately return None.

A focused executable check exercised serial and threaded execution, straggler resubmission, and comparison-result handling. The updated code reports completed None tasks correctly, accepts a resubmitted None result as completion before a late duplicate can replace it, and avoids indexing None in comparison mode. The focused parallelizer suite passed with 14 tests.

Confidence Score: 5/5

T-Rex Logs

What T-Rex did

Ran a focused executable probe against the parent revision to validate the handling of None results and observed the old behavior where None results appeared as 0/3 and the late original execution completed after 1.351 seconds.
Ran the same probe after the change and observed the improved behavior: serial and threaded runs show 3/3, the resubmitted None completes the straggler at 1.002 seconds, and compare-results runs complete without indexing None.
Executed uv run --locked pytest -q tests/utils/test_parallelizer.py and confirmed all 14 focused tests passed.
Authored a regression probe source at trex-artifacts/parallelizer-none-regression.py to exercise serial, threaded, straggler, and comparison flows.
Collected and inspected the before- and after-output logs and suite results to document the post-PR behavior, including pr10142-none-results-01-before.log, pr10142-none-results-02-after.log, and pr10142-parallelizer-suite.log.

_{Ran code and verified through T-Rex}

_{Reviews (1): Last reviewed commit: "fix(dspy): track completed None results ..." | Re-trigger Greptile}

fix(dspy): track completed None results in ParallelExecutor

bef0878

Co-authored-by: Isaac Miller <isaacbmiller@users.noreply.github.com> Amp-Thread-ID: https://ampcode.com/threads/T-019fcd6b-0068-752e-b9ce-2fd40b7f061b Co-authored-by: Isaac Miller <isaac@cmpnd.ai>

isaacbmiller merged commit 485ce44 into main Aug 5, 2026
15 checks passed

codechrl mentioned this pull request Aug 9, 2026

fix(utils): don't silently swallow harness-level exceptions in ParallelExecutor #10189

Open

Navigation Menu

Uh oh!

fix(dspy): track completed None results in ParallelExecutor - #10142