From 6ca82ed7c9bef9ee4c0df55bb7ab927c3e84ee33 Mon Sep 17 00:00:00 2001 From: Luke Curley Date: Thu, 12 Nov 2026 11:05:13 -0800 Subject: [PATCH 1/5] chore(quest): claim quest/m0/checks/README.md Quest-Claim: c503136cfbc6b9a6 From b5dff7553e58b86aa043735a4b93270fa332372a Mon Sep 17 00:00:00 2001 From: Luke Curley Date: Thu, 12 Nov 2026 11:05:56 -0800 Subject: [PATCH 2/5] chore(quest): record the first warm queue build's cache stats Co-Authored-By: Claude Opus 5.5 --- quest/m0/checks/README.md | 22 ++++++++++++++++++++-- 1 file changed, 20 insertions(+), 2 deletions(-) diff --git a/quest/m0/checks/README.md b/quest/m0/checks/README.md index 3faac112c..a351ffcc0 100644 --- a/quest/m0/checks/README.md +++ b/quest/m0/checks/README.md @@ -15,6 +15,24 @@ stays the rule, and costs less now that it selects by the diff ([#421](https://github.com/moq-dev/rust2/pull/421)); CONTRIBUTING.md and the skills keep their rules. All of it is m0, since every agent pays the cost now. -Its children have landed; what remains is the measurement: once a queue build logs -`mbx[cache]: object cache: ...`, confirming the warm gate, report the gate's and local CPU +Its children have landed; what remains is the measurement: report the gate's and local CPU against the baseline above, cold and warm, and record any gap as a new quest. +The CPU numbers come from the runner host's systemd accounting, which an agent without +access to nixos2 cannot read; queue wall time from the Actions API stands in until then. + +The first warm queue build that compiled, +[run 43265518031](https://github.com/moq-dev/rust2/actions/runs/43265518031) for +[#715](https://github.com/moq-dev/rust2/pull/715), logged `mbx[cache]` with 0 hits, +55 misses, and 10 bypassed (`unsupported-crate-type`), storing 66.2 MiB: it seeded an empty +store, so it shows the store fills, not that it serves. Its Build took 7m18s, against 9m23s +to 23m46s for the three queue builds before it. The only later queue build, +[run 43286222929](https://github.com/moq-dev/rust2/actions/runs/43286222929) for #718, +compiled nothing (23 s, diff-selected checks only). The measurement starts at the next +queue build that compiles Rust: if its `object cache` line still shows 0 hits, the warm +gate does not work and that is the gap quest; otherwise compare its CPU with the baseline. + +## Waiting + +- Maintainer: read the queue builds' CPU from nixos2's systemd accounting (the baseline's + source), cold and warm, since agents cannot reach the host; done when this Plan records + them. Considered: Actions timing alone, which gives wall time, not CPU. From 6a8a3dc8088df6eaad362a4f29ac57cf9a0c8547 Mon Sep 17 00:00:00 2001 From: Luke Curley Date: Thu, 12 Nov 2026 15:08:10 -0800 Subject: [PATCH 3/5] chore(quest): finish m0/checks on its gate measurement, add the object cache miss Co-Authored-By: Claude Opus 5.5 --- quest/m0/checks/README.md | 38 ----------------------------------- quest/m0/object-cache-miss.md | 22 ++++++++++++++++++++ 2 files changed, 22 insertions(+), 38 deletions(-) delete mode 100644 quest/m0/checks/README.md create mode 100644 quest/m0/object-cache-miss.md diff --git a/quest/m0/checks/README.md b/quest/m0/checks/README.md deleted file mode 100644 index a351ffcc0..000000000 --- a/quest/m0/checks/README.md +++ /dev/null @@ -1,38 +0,0 @@ -# [S] Checks at moq's cost - -## Goal - -The merge queue's build and agents' local checks cost what moq's do: `just ci` runs only what a -change touches, builds warm, and runs the differential fixtures optimized, while a nightly -`just gate` runs the full suite. Measured against the 2026-11-01 audit's baseline: about -33 CPU-hours a day in the gate (65 CPU-minutes a run) and 100 to 130 locally. - -## Plan - -Decided with the maintainer on 2026-11-01, from an audit of the gate's and agents' checks -(nixos2 and the host's systemd accounting): copy how moq works. `just ci` before every push -stays the rule, and costs less now that it selects by the diff -([#421](https://github.com/moq-dev/rust2/pull/421)); CONTRIBUTING.md and the skills -keep their rules. All of it is m0, since every agent pays the cost now. - -Its children have landed; what remains is the measurement: report the gate's and local CPU -against the baseline above, cold and warm, and record any gap as a new quest. -The CPU numbers come from the runner host's systemd accounting, which an agent without -access to nixos2 cannot read; queue wall time from the Actions API stands in until then. - -The first warm queue build that compiled, -[run 43265518031](https://github.com/moq-dev/rust2/actions/runs/43265518031) for -[#715](https://github.com/moq-dev/rust2/pull/715), logged `mbx[cache]` with 0 hits, -55 misses, and 10 bypassed (`unsupported-crate-type`), storing 66.2 MiB: it seeded an empty -store, so it shows the store fills, not that it serves. Its Build took 7m18s, against 9m23s -to 23m46s for the three queue builds before it. The only later queue build, -[run 43286222929](https://github.com/moq-dev/rust2/actions/runs/43286222929) for #718, -compiled nothing (23 s, diff-selected checks only). The measurement starts at the next -queue build that compiles Rust: if its `object cache` line still shows 0 hits, the warm -gate does not work and that is the gap quest; otherwise compare its CPU with the baseline. - -## Waiting - -- Maintainer: read the queue builds' CPU from nixos2's systemd accounting (the baseline's - source), cold and warm, since agents cannot reach the host; done when this Plan records - them. Considered: Actions timing alone, which gives wall time, not CPU. diff --git a/quest/m0/object-cache-miss.md b/quest/m0/object-cache-miss.md new file mode 100644 index 000000000..32d63a3c5 --- /dev/null +++ b/quest/m0/object-cache-miss.md @@ -0,0 +1,22 @@ +# [S] Object cache hits on an identical queue rebuild + +## Goal + +A merge-queue build that recompiles crates an earlier queue build already compiled serves +them from mbx's object cache, so the `mbx[cache]: object cache` line shows hits for them. + +## Plan + +Two queue builds on the `nixos-rust2-ci` runner (nspawn container `rust2-ci` on `nixos`) +compiled the same 55 crates and both logged 0 hits, 55 misses, 66.2 MiB stored locally: +[run 43265518031](https://github.com/moq-dev/rust2/actions/runs/43265518031) and +[run 43290025781](https://github.com/moq-dev/rust2/actions/runs/43290025781). A later one, +[run 43304642908](https://github.com/moq-dev/rust2/actions/runs/43304642908), got 17 hits +of 204 lookups. Each job still finished in 9 to 19 CPU-minutes against the 65 of the +2026-11-01 baseline, so the cache is headroom, not the gate's main saving. + +Find the cause before fixing it: whether the local store persists across queue jobs in the +container (path, owner, cleanup between jobs), and whether keys change between identical +runs (paths, env, timestamps in the key). Reproduce by running the same build twice in the +container and diffing the keys mbx logs. Fix at the root, and confirm with a queue build +whose repeated crates hit. From 8d29ea6acd12cfc41151cbc1af0e560ebc8c4d39 Mon Sep 17 00:00:00 2001 From: Luke Curley Date: Thu, 12 Nov 2026 15:18:36 -0800 Subject: [PATCH 4/5] chore(quest): source the queue CPU figures Co-Authored-By: Claude Opus 5.5 --- quest/m0/object-cache-miss.md | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/quest/m0/object-cache-miss.md b/quest/m0/object-cache-miss.md index 32d63a3c5..c7501f68b 100644 --- a/quest/m0/object-cache-miss.md +++ b/quest/m0/object-cache-miss.md @@ -12,8 +12,9 @@ compiled the same 55 crates and both logged 0 hits, 55 misses, 66.2 MiB stored l [run 43265518031](https://github.com/moq-dev/rust2/actions/runs/43265518031) and [run 43290025781](https://github.com/moq-dev/rust2/actions/runs/43290025781). A later one, [run 43304642908](https://github.com/moq-dev/rust2/actions/runs/43304642908), got 17 hits -of 204 lookups. Each job still finished in 9 to 19 CPU-minutes against the 65 of the -2026-11-01 baseline, so the cache is headroom, not the gate's main saving. +of 204 lookups. Each job still used 9 to 19 CPU-minutes (the `quest-gate` scopes' `Consumed` +CPU in the container journal, `journalctl -D /var/lib/nixos-containers/rust2-ci/var/log/journal` +on `nixos`, not Actions wall time) against the 65 of the 2026-11-01 baseline, so the cache is headroom, not the gate's main saving. Find the cause before fixing it: whether the local store persists across queue jobs in the container (path, owner, cleanup between jobs), and whether keys change between identical From d305dc3b7ea6b86e70937ff8b5b9ca3c044ae7f6 Mon Sep 17 00:00:00 2001 From: Luke Curley Date: Thu, 12 Nov 2026 15:23:39 -0800 Subject: [PATCH 5/5] chore(quest): record keeping the object-cache-miss follow-up Co-Authored-By: Claude Opus 5.5 --- quest/m0/object-cache-miss.md | 3 +++ 1 file changed, 3 insertions(+) diff --git a/quest/m0/object-cache-miss.md b/quest/m0/object-cache-miss.md index c7501f68b..608d30bb1 100644 --- a/quest/m0/object-cache-miss.md +++ b/quest/m0/object-cache-miss.md @@ -21,3 +21,6 @@ container (path, owner, cleanup between jobs), and whether keys change between i runs (paths, env, timestamps in the key). Reproduce by running the same build twice in the container and diffing the keys mbx logs. Fix at the root, and confirm with a queue build whose repeated crates hit. + +Decided 2026-11-12 (maintainer, the pick of [#720](https://github.com/moq-dev/rust2/pull/720)): +✅ keep this quest (rejected: drop it, since gate CPU already beats the baseline).