PkgEval
Check installation and package tests of an exact registered release on Julia stable and nightly. PkgEval is an external Linux sandbox tool, used by the optional Registered PkgEval workflow; it is not a PerfChecker performance collector.
To evaluate an installed release, open Actions → Registered PkgEval → Run workflow and select source_kind=registered, its package_version, a general_revision and optionally the expected registered tree. The selector resolves immutable source identity before evaluation. Keep stable and nightly verdicts separate.
The completed Core 1.1.0 run tested registered tree 753e41cad1c984640a2337b314b70ad6932626ad (release commit 4e81603c57b47bc08372992492fdf2cbbd718b3b). Its recorded result is:
| Runtime | Upstream status | Package assertions | Configuration |
|---|---|---|---|
| Julia 1.13.1 stable | test | 2,380 PASS; 1 Broken | Full suite; shared cache disabled; upstream compilation policy |
| Julia 1.14.0-DEV.3606 nightly | test | 2,380 PASS; 1 Broken | Same policy, evaluated separately |
Both used a configured 90-minute sandbox budget. That is a limit, not measured duration. These results establish the selected release's package-test outcome on those runtimes; they do not measure speedups or qualify all interfaces/hardware counters.
Read artifacts and choose another source
Download pkgeval-request, pkgeval-stable and pkgeval-nightly. Match request source kind, version/tree, General/controller revisions and workflow identity. Inspect each result.toml, evaluation.log and PkgEval-Manifest.toml, including status/reason and phase-completion fields. A successful package-test status is test. Test-phase duration_seconds is CPU time in the pinned controller; the final “PkgEval … after …s” log line is total elapsed sandbox wall time. Missing/zero unavailable timing fields must not become zero-cost claims.
The completed run's stable artifact 11707536498 has SHA-256 2e05e950999151f895bc11a3b00ec3a251a609ceefb76b7b50f28ad7717cf04b; nightly artifact 11707694207 has SHA-256 ca6fd33ed193f61330379540b6eb7046b9a024442b4e995c23bf41530e7da1a7. The original receipts remain unchanged.
For an unpublished candidate use source_kind=git_candidate, a lowercase full 40-character candidate_revision, and a General revision for dependencies; leave version/expected-tree fields empty. Candidate records identify repository/revision and do not claim registered-tree verification. A candidate pass cannot repair an immutable registered archive.
Manual defaults are package_images=upstream and time_limit_minutes=45; selecting 90 gives 5,400 seconds for the complete sandbox evaluation and a separate 120-minute job limit. Setup, installation, precompilation and testing share the sandbox budget. package_images=yes changes compilation policy; do not equate that outcome with upstream-default qualification. Scheduled source selection/configuration is recorded in the workflow request.
Earlier 45-minute attempts remain failures: stable reached kill/time_limit without an observed failed assertion, while nightly failed downloading Julia (code 28) before tests. The older 1.0.1 evaluation remains its own failed archive result. A later pass does not rewrite any of these receipts. Observer cleanup_complete describes observer-owned resources, not a guarantee that every process on the host stopped.
Local use requires PkgEval's Linux/kernel/container prerequisites. See its setup instructions.