mirror of
https://github.com/windmill-labs/windmill.git
synced 2026-08-19 16:02:14 +00:00
Ruben Fiszel
045d12043e
* [ee] feat(queue): duration-weighted fairness admission atomic Add the `WORKSPACE_FAIRNESS_ADMISSION_PPM` atomic that the EE `workspace_fairness_ee::refresh_overloaded` writes on each refresh (see companion EE PR). The atomic is read on every pull by `should_admit_capped` to decide whether the dispatch goes down the standard or fairness path. Defaults to 10_000 (= admit all) so the pre-fairness behaviour is preserved until the first refresh fires. OSS stub in `workspace_fairness.rs` continues to return `true` unconditionally, so non-EE builds are bit-identical. * docs(queue): consolidate full fairness algorithm into workspace_fairness.rs Move the algorithm doc — what "overloaded" means in worker-seconds, the duration-weighted admission derivation, coordinated refresh structure, audit emission, the SQL perf constraints (no params CTE, drive running side from v2_job_runtime), and EE gating — into the OSS surface module where it is readable without EE access. The EE file becomes implementation only. Also bump ee-repo-ref to the EE commit that strips the duplicate doc. * docs(queue): clarify ADMISSION_PPM default is "admit all", not count-based Addresses CI review (claude[bot]): the `10_000` initial value is the "admit all" no-op default that applies before the first refresh classifies an overloaded set — not the count-based value (which would be `target * 10_000`). The count-based form is the empty-bucket fallback inside `compute_admission_ppm`, a different thing. * chore(queue): point ee-repo-ref at EE main (fairness admission merged via #593) * fix(queue): duration-weighted admission uses unclamped service-time window Bumps ee-repo-ref to the EE fix (windmill-ee-private#596) that sources D_c/D_u for the admission probability from a separate 60s service-time window of true `duration_ms`, instead of the occupancy aggregation whose per-job contributions are clamped to the 10s occupancy window. The clamp truncated D_c for capped jobs longer than the window, under-admitting the duration skew (true 34s jobs → ~86% effective share instead of the target 65%). Occupancy worker-seconds still drive overload classification. Updates the algorithm doc in workspace_fairness.rs accordingly. Note: ee-repo-ref points at the EE feature branch; re-point to EE main once #596 merges.