diff --git a/quest/m0/ci/mbx-r2-verify.md b/quest/m0/ci/mbx-r2-verify.md index 8ddba0b2e73..a3db7d534a1 100644 --- a/quest/m0/ci/mbx-r2-verify.md +++ b/quest/m0/ci/mbx-r2-verify.md @@ -307,13 +307,29 @@ alignment`, `gh test-integration`, `test qmux`, `test drill-sensitivity`, The sums above leave out the rows marked new or split and `Shipping sizes`, which have no comparable baseline. +Evidence checked 2026-11-14, scheduled run +[43601796176](https://github.com/moq-dev/moq/actions/runs/43601796176) +(11-14 13:50 UTC, the first after #8631, mbx 1.30.0 at 20 GiB). The +[Windows workspace build](https://github.com/moq-dev/moq/actions/runs/43601796176/job/132468015187) +had 621 hits, 156 misses, 390 not looked up and 499 prefetched. It +downloaded 13.5 GiB and uploaded 600.9 MiB, with no `download budget exceeded` +warnings, against 92 / 1,062 / 13 at the 10 GiB cap on 11-13. Once the +budget stopped binding, Windows hit 80% of the keys it looked up, so the +budget caused its misses. `mbx analyze` still lists all 156 misses as +"misses with no earlier recording". The 390 not looked up, up from 13, and +670 `NoSuchKey` lines in the job log are still the expiry wave, so this +night is not a clean baseline. Posted to +[jdx/mr-boxington#950](https://github.com/jdx/mr-boxington/discussions/950#discussioncomment-19616412) +(Windows recovering once the budget stops binding, and 13.5 GiB already +over the 10 GiB default that #951's half-budget prefetch cap splits). The +drafted issue below was not filed and is obsolete. + What the after-#8631 nights must show here, read from the same summaries the [failed-puts quest](/quest/m0/ci/mbx-failed-puts.md) already collects for its own after-measurement: -- Windows: whether its workspace build hits most looked-up keys once its - download stays under 20 GiB with no budget refusals. The first such - night settles the upstream report (Waiting). +- Windows: settled on 11-14 (above). Later nights only need to confirm + that the hits hold once the expiry wave passes. - Linux: whether full history now raises each job's hits above its depth-1 range. If the jobs still sit inside it once restores stop failing, plan a quest that returns them to `fetch-depth: 1`, since full history then costs @@ -350,23 +366,6 @@ placeholder with the first night after #8631. ## Waiting -- GitHub Actions: the first scheduled Nightly after #8631 (expected - 2026-11-14) settles the Windows upstream report, by its Windows workspace - build's `mbx[cache]: object cache` line and any `download budget exceeded` - warnings. Record that night's numbers in the Plan, then: - - If Windows hits most looked-up keys, or still misses with budget - refusals, comment on - [jdx/mr-boxington#950](https://github.com/jdx/mr-boxington/discussions/950) - only if the numbers add something it lacks, such as Windows hits - recovering once the budget stops binding, or refusals at 20 GiB. Skip it - otherwise, and file no issue. - - If Windows still misses most looked-up keys with no budget refusals, file - the drafted issue (Plan), its placeholder filled from that night. - - The maintainer approved both posts. Each follows mr-boxington's - CONTRIBUTING.md and AGENTS.md AI-disclosure rules and ends with the - `(Written by with )` line. Link what was posted, or say - nothing was, in the Plan, then drop this item. - moq-dev/moq: no earlier than 2026-11-26, an agent on the maintainer's desktop collects `mbx-moq-dev`'s size and request counts for the first month and checks that the six old oversized manifests expired. Size comes