Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 5 additions & 1 deletion quest/m0/ci/hakari-followups.md
Original file line number Diff line number Diff line change
Expand Up @@ -82,4 +82,8 @@ Public API: none. Wire: none.
baseline and 2026-11-06 numbers in the Plan. If the cache
cost is material, present the measured tradeoff to the maintainer; recommend
keeping the narrow final exclusions unless the measurements justify another
design.
design. That job timed out on five of the six scheduled runs from
2026-11-05 to 11-10 and on every one since 11-07, so a complete run waits on
[it fitting its hour](/quest/m0/ci/nightly-rust-test-hour.md). That change
regroups the job's builds, so compare only the mbx summaries of builds whose
crates, features, and profile match the baseline's.
31 changes: 31 additions & 0 deletions quest/m0/ci/nightly-relay-session-hour.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,31 @@
# [S] Nightly's relay session benchmarks fit their hour

## Goal

`Nightly (Relay session benchmarks)` (`just gh nightly rust-relay`, a paired
group that measures this commit and the previous Nightly's) finishes inside
its 60-minute `timeout-minutes` on a scheduled run and keeps its paired
regression gate.

## Plan

The job was cancelled at the timeout on the 2026-11-08, 11-09, and 11-10
scheduled runs. Unlike the other cancelled jobs it does not hang:
[the 11-10 run](https://github.com/moq-dev/moq/actions/runs/42842827718/job/129922278379)
was still finishing `relay_session_join_broadcasts` cells seconds before the
cancel. The scheduled runs from 11-01 to 11-05 failed inside the workload after
14 to 20 minutes, the
[11-01 job](https://github.com/moq-dev/moq/actions/runs/41313466339/job/124868517800)
with `Evicted`. The
[11-06](https://github.com/moq-dev/moq/actions/runs/42236107954) and
[11-07](https://github.com/moq-dev/moq/actions/runs/42377867523) runs
completed the sweep in about 25 and 23 minutes (their jobs failed later, at
`Fail on paired regressions`). Find what grew between those sweeps and the
60-minute cancellations from 11-08, then trim it or split the group across
matrix entries, as
[the Rust benchmark tests quest](/quest/m0/ci/nightly-rust-test-hour.md) does
for its job. The sweep still covers both fan-out axes.

## Related

- [Rust benchmark tests fit their hour](/quest/m0/ci/nightly-rust-test-hour.md) - the same timeout, caused by builds
34 changes: 34 additions & 0 deletions quest/m0/ci/nightly-rust-test-hour.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,34 @@
# [M] Nightly's Rust benchmark tests fit their hour

## Goal

`Nightly (Rust benchmark tests)` (`just gh nightly rust-test`) finishes well
inside its 60-minute `timeout-minutes` on a scheduled run, with every benchmark
it runs today still measured in a release build, one at a time, and reported.
Splitting it across jobs and cutting what it builds are both in scope.

## Plan

Decided 2026-11-10 (maintainer, /quest-unblock): ✅ split the job and reduce
its size (rejected: finish the hakari measurement with one run; raise the
timeout).

Evidence: the job was cancelled at the timeout on the 2026-11-03, 11-05,
11-07, 11-08, 11-09, and 11-10 scheduled runs; the 11-04 and 11-06 runs
finished in 57 to 59 minutes, and the last to pass, on 11-02, took 34. In
[the 11-10 run](https://github.com/moq-dev/moq/actions/runs/42842827718/job/129922278502)
the benchmarks themselves ran for about two minutes in all. The hour went to
`sh/gh/nightly.sh`'s `rust-test` group compiling a separate release build per
`nextest` invocation (3 to 5.5 minutes each, about ten of them), and the job
was still compiling `moq-ffi` when it was cancelled.

Start from the builds: fewer `nextest` invocations sharing one release build
where the feature sets allow it, then split what remains across matrix entries
so no entry nears the hour. Keep `sh/gh/nightly.test.py`'s guarantees (release
builds, one JUnit report per run, the report publish step) or update them with
the change.

## Related
Comment thread
pullfrog[bot] marked this conversation as resolved.
Resolved

- [hakari follow-ups](/quest/m0/ci/hakari-followups.md) - its mbx measurement needs a complete run of this job
- [Relay session benchmarks fit their hour](/quest/m0/ci/nightly-relay-session-hour.md) - the same timeout in another Nightly job
25 changes: 25 additions & 0 deletions quest/m0/js-subscribe-drop-hang.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,25 @@
# [S] The JS subscribe-drop benchmark finishes

## Goal

`just js bench subscribe-drop` (`js/net/bench/subscribe-drop.ts`) finishes and
prints its rows, and the root cause of its hang is fixed, so
`Nightly (JS benchmarks)` finishes again.

## Plan

`Nightly (JS benchmarks)` was cancelled at its 60-minute timeout on every
scheduled run from 2026-11-04 to 11-10; it last passed on 11-02, in about
eight minutes. In
[the 11-10 run](https://github.com/moq-dev/moq/actions/runs/42842827718/job/129922278808)
`subscribe-drop.ts` printed its CSV header
(`publishers,subscribers,groups,drop_us`) and nothing for the job's last 57
minutes, and the benchmarks after it never ran. The
[11-04 run](https://github.com/moq-dev/moq/actions/runs/41852517544/job/126652296690)
stalled at the same header. Reproduce it locally; if the
hang is in `@moq/net` rather than the benchmark, a test in the merge gate
covers it.

## Related

- [session_delivery_compression finishes on lite-07](/quest/m0/lite07-compression-bench-hang.md) - another Nightly benchmark that hangs
26 changes: 26 additions & 0 deletions quest/m0/lite07-compression-bench-hang.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,26 @@
# [M] session_delivery_compression finishes on moq-lite-07

## Goal

The `session_delivery_compression` Criterion group (moq-net's `session`
bench target, `rs/moq-net/benches/session/mod.rs`) finishes every
`moq-lite-07-wip` cell, and the root cause of the hang it shows is fixed. If
the hang is in moq-net rather than the benchmark harness, a test reproduces it
in the merge gate.

## Plan

`Nightly (Rust bench targets)` was cancelled at its 60-minute timeout on the
2026-11-02, 11-05, 11-06, 11-09, and 11-10 scheduled runs, and failed without
timing out on 11-03, 11-04, 11-07, and 11-08.
In [the 11-10 run](https://github.com/moq-dev/moq/actions/runs/42842827718/job/129922278472)
every `moq-lite-06` cell finished, then
`session_delivery_compression/moq-lite-07-wip/relays=1/publishers=1/broadcasts=1/viewers=1`
printed `Benchmarking` and nothing for the job's last 34 minutes, so the hang
starts with the first lite-07 cell, and every later bench in the job goes
unmeasured. Reproduce it locally, bisect against the lite-07 changes
if it is a regression, and fix it at the root.

## Related

- [JS subscribe-drop finishes](/quest/m0/js-subscribe-drop-hang.md) - another Nightly benchmark that hangs
5 changes: 2 additions & 3 deletions quest/m2/promote-cluster-routes.md
Original file line number Diff line number Diff line change
Expand Up @@ -24,8 +24,7 @@ no new quest under the old path first.
## Waiting

- moq-dev/moq: the open PRs under `quest/m2/cluster-routing/routes/` land or
close ([#8401](https://github.com/moq-dev/moq/pull/8401),
[#8397](https://github.com/moq-dev/moq/pull/8397) and
[#8375](https://github.com/moq-dev/moq/pull/8375) on 2026-11-09). Check
close (only [#8443](https://github.com/moq-dev/moq/pull/8443) on
2026-11-10). Check
with `gh pr list -R moq-dev/moq --search "head:quest/m2/cluster-routing/routes/"`
printing nothing.
34 changes: 18 additions & 16 deletions quest/m3/capture-list.md
Original file line number Diff line number Diff line change
Expand Up @@ -115,15 +115,29 @@ as given:
- **Placement.** A window moving or resizing is a change like a retitle, so
the watch's list stays current; coalesce a drag so it does not wake the
caller per frame. X11 has `ConfigureNotify`, Windows `EVENT_OBJECT_LOCATIONCHANGE`,
and macOS reports bounds in each listing, with the same caveat as the
Waiting item for events.
and macOS reports bounds in each listing, noticed the same way as the
other window changes.
- **Changes.** X11 has `PropertyNotify` on the root's client list and each
window's title, and RandR for monitors, which `capture/x11.rs` already
selects. Cameras have udev and PipeWire's registry on Linux, device
notifications on Windows, and AVFoundation's connect and disconnect
notifications on macOS. Windows has `SetWinEventHook`. macOS reports
another app's windows only through the Accessibility API, which is the
decision under Waiting.
another app's windows only through the Accessibility API, so the caller
picks a macOS mode when it opens the watch, as decided below.
- **macOS window changes.** Decided 2026-11-10 (maintainer,
/quest-unblock): ✅ both modes, the application picks one at setup
(rejected: re-list only; Accessibility only; app launch, quit, and
activation only). In the Accessibility mode moq-video observes each app's
window events, and setup refuses with `Unsupported` naming the missing
permission rather than falling back. In the basic mode, which needs no
Accessibility permission, moq-video re-lists windows at a documented rate only while a
caller holds a watch, and wakes it on a difference; no caller polls.
Re-listing is wasteful, so it never runs without a watch. Window pictures
still need Screen Recording permission in either mode, as any
ScreenCaptureKit capture does. The Mac tests
cover both modes.
[The foreground app](/quest/m4/capture-foreground.md) takes the same mode,
so one permission serves both.

Tests, sized to what each runner has:

Expand All @@ -147,18 +161,6 @@ Tests, sized to what each runner has:
Public API: breaking for `capture::{Camera, Display, Window, App}` and the
listing calls. Wire: none.

## Waiting

- Maintainer: how macOS notices another app's windows opening, closing, and
being retitled, which only the Accessibility API reports as events. An
agent cannot choose it because one option adds a permission prompt to every
embedder. Options: (1, recommended) moq-video re-lists inside the watch at
a stated rate while a caller holds one, on macOS windows only, so no caller
polls and no permission is added; (2) ask for the Accessibility permission
and observe each app; (3) notify on app launch, quit, and activation only,
missing a window that opens inside a running app. Done when this item names
the pick.

## Related

- [macOS capture listing futures are Send](/quest/m3/capture-shareable-send.md) - the ScreenCaptureKit listing this rewrites; whatever replaces it stays `Send`
Expand Down
31 changes: 13 additions & 18 deletions quest/m4/capture-foreground.md
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,9 @@ It works on macOS, Windows, and X11 under a window manager that publishes
the active window. Where the desktop does not tell an app, which is every
Wayland session, setting it up is an `Unsupported` error naming the reason,
decided before the first report, never a report that stays empty or stale.
On macOS the application picks at setup between following every window,
which needs the Accessibility permission, and following the app with no
permission, which misses a switch between two windows of one app.

## Plan

Expand Down Expand Up @@ -44,7 +47,16 @@ Recommendations, for the PR to confirm against each system:
changes. Those notifications arrive only while the main run loop runs, so
setup either checks for one or says what the caller must be running; a
watch that never fires is the failure to avoid. Which window of that app
is in front is the decision under Waiting.
is in front follows the caller's macOS mode, the same one
[the capture listing](/quest/m3/capture-list.md) takes. Decided 2026-11-10
(maintainer, /quest-unblock): ✅ both modes, the application picks one at
setup (rejected: app by event only; Accessibility only; app only). The
Accessibility mode follows the focused window, including a switch between
two windows of one app, and setup refuses with `Unsupported` naming the
missing permission rather than falling back. The basic mode needs no
permission and reports the app by event and its front window as of that
event, so a window switch inside one app is not noticed until the app
changes; the docs say so. The Mac test covers both.
- **Wayland.** Refused even with XWayland running. The compositor tells no
client which window has focus, and XWayland's `_NET_ACTIVE_WINDOW` covers
X11 clients only, so it would name a stale app, or none, whenever a native
Expand All @@ -68,23 +80,6 @@ Tests:

Public API: additive. Wire: none.

## Waiting

- Maintainer: which window macOS reports as in front. The window server
orders windows front to back without a permission, but only the
Accessibility API posts when focus moves between two windows of one app.
An agent cannot choose it because one option adds a permission prompt to
every embedder. Options: (1, recommended) report the app by event and its
front window as of that event, and narrow the Goal, the change contract,
and the Mac test to say a window switch inside one app is not noticed on
macOS. A caller can capture the whole app with `Source::App`, but
`open_app` captures the main display only, so a game on another display is
missed until the app capture follows the app's display; (2) ask for the
Accessibility permission and follow the focused window, keeping the Goal
as written; (3) report the app only on macOS, narrowing the Goal the same
way as (1). Done when
this item names the pick.

## Required

- [A capture listing a picker can draw](/quest/m3/capture-list.md) - the app and window identities this reports, on every desktop
Expand Down
Loading