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
2 changes: 2 additions & 0 deletions pyproject.toml
Original file line number Diff line number Diff line change
Expand Up @@ -50,6 +50,8 @@ exclude = [
{ path = "**/target/**", format = "sdist" },
# Media fixtures are only read by Rust tests, not source installs.
{ path = "**/test_data/**", format = "sdist" },
# The vendored WebRTC stack's RTP captures, which its crate also excludes.
{ path = "**/tests/data/**", format = "sdist" },
]

# Listed explicitly (not `py/*`) so build output like py/dist isn't mistaken
Expand Down
74 changes: 74 additions & 0 deletions quest/m1/js-bundle-audit.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,74 @@
# [M] Audit what the JS bundles carry

## Goal

Know what each consumer bundle the Nightly size report measures (`net`, `hang`,
`watch`, `watch-ui`, and `publish`, built by `sh/size/bundles.ts`) carries, and
shrink each by what can be split out, lazy-loaded, or dropped without changing
a package's behavior. Then Nightly's 30-day size comparison passes again,
either because the bundles shrank or because a post-growth report became its
anchor.

## Plan

Decided 2026-11-13 (maintainer, /quest-complete, in
[moq-dev/moq#8654](https://github.com/moq-dev/moq/pull/8654)): ✅ accept the JS
bundle growth as feature growth (lite-07 routes, the byte pool, Clock types)
and plan this audit (rejected: accepting it without an audit; holding until
the bundles shrink). The audit is not a gate on the acceptance.

Measured 2026-11-13 on `main` at `a026b71b1`, built as Nightly does, against the
retained reports. After
[moq-dev/moq#8671](https://github.com/moq-dev/moq/pull/8671) dropped
`sorted-btree` and the test-only `Manual` clock from every bundle, brotli is
2% to 3% below the 11-13 report and still 9% to 12% above the 11-07 one:

| Bundle | 10-22 | 11-07 | 11-13 | `main` |
|---|---:|---:|---:|---:|
| `net` | 98,909 | 136,880 | 157,282 | 153,149 |
| `hang` | 119,937 | 157,584 | 178,277 | 174,142 |
| `watch` | 133,560 | 175,294 | 196,114 | 192,662 |
| `watch-ui` | 144,409 | 186,072 | 206,966 | 203,518 |
| `publish` | 131,060 | 169,777 | 189,804 | 186,314 |

Where the 30-day comparison (10%) stands, since the failure is the anchor, not
a new regression:

- It starts failing on 2026-11-21 against the 10-22 report, where `main` is
1.41x to 1.55x. The anchor then advances to the newest retained report at
least 30 days old, and every report before 11-07 is at least 10% below
`main` for every bundle.
- From 12-07 the 11-07 report is the anchor. `net` (1.119x) and `hang`
(1.105x) still fail by 1.7% and 0.5% of their size; `watch`, `watch-ui`, and
`publish` pass. This is the stretch the audit can clear.
- From 12-13 the 11-13 report is the anchor, and `main` passes unless it grows
again.
- Before 12-07 the anchors are the 10-22 to 11-06 reports, which `main`
exceeds by 13% (11-06) to 55% (10-22) for `net`. No non-breaking shrink
clears 11-21 to 11-30 (1.47x to 1.55x), so those nights are red by design,
and [Nightly artifact sizes](/sh/size/README.md) says the comparison is
neither waived nor reset (a reset deletes every series). Do not read them as
a new regression, and still read each night's other lines.

Starting points, from the static chunks of the `main` metafiles (the lazy
chunks, such as the IETF stack and the libav polyfill, are already outside
them):

- `net` is about 577 KB raw, 446 KB of it `js/net/dist`: `track.js` 56 KB,
`lite/subscriber.js` 45 KB, `lite/publisher.js` 30 KB, `origin.js` 30 KB,
`resume.js` 26 KB, `group.js` 18 KB. Then `pako` 42 KB, `js/qmux` 44 KB, and
`js/pattern` 14 KB. Does `pako` earn its size against `CompressionStream`?
- `hang`, `watch-ui`, and `publish` carry the same `net`, plus `zod` (37 to
39 KB, against 9 KB in `net`), so what in `js/hang` pulls the larger part of
`zod` into every consumer is the first question.
- `watch-ui` adds 123 KB of `js/watch/dist` and `publish` 71 KB of
`js/publish/dist`; each is a bundle of its own to explain.

Per package, record in the PR: what the bundle carries (the top inputs by
bytes), what can be split or lazy-loaded and what it costs a consumer, and what
the audit changed with before and after byte counts from `bundles.ts`. A change
that needs a breaking release for a published consumer, such as moving an export
behind a subpath, is its own m2 quest; the audit lands the non-breaking ones.
The thresholds stay as they are.

Public API: none unless a finding says so. Wire: none.
3 changes: 2 additions & 1 deletion sh/size/report.py
Original file line number Diff line number Diff line change
Expand Up @@ -97,7 +97,8 @@ def main():
(args.directory / "summary.md").write_text(summary)
print(summary)
if failures:
raise SystemExit(1)
# Repeated last, so the failing step's log ends with why rather than the build-time table.
raise SystemExit("size regression:\n" + "\n".join(failures))


if __name__ == "__main__":
Expand Down
13 changes: 13 additions & 0 deletions sh/size/report.test.py
Original file line number Diff line number Diff line change
Expand Up @@ -125,7 +125,20 @@ def run(current):
previous = row("2026-10-10", 106) | {"revision": "a" * 40}
failed, retained = run(previous)
self.assertEqual(failed.returncode, 1, failed.stderr)
# The log ends with the regression, after the build-time table.
regression = "artifact: 106 > 100 bytes + 5% (previous: 2026-10-09)\n"
self.assertTrue(failed.stderr.endswith(regression), failed.stderr)
self.assertEqual(retained[-1], previous)
# Through one pipe, as the job's log is, the regression still comes last.
earlier = root / "earlier.json"
earlier.write_text(json.dumps([row("2026-09-10"), row("2026-10-09")]))
merged = subprocess.run(
[sys.executable, str(Path(report.__file__)), str(root), "--history", str(earlier)],
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
text=True,
)
self.assertTrue(merged.stdout.endswith(regression), merged.stdout)

# A stable next report compares against the failed point, not the last success.
current = row("2026-10-11", 106) | {"revision": "b" * 40}
Expand Down