You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
perf(test): run the API unit coverage lane at four workers - #1480
Run API unit coverage at four isolated workers per CI shard now that Bun 1.4.3 has removed the known epoll_ctl abort path. Keep one bun test invocation and one LCOV per unit shard.
Changes
Set apps/api's test:coverage:unit to --parallel=4 and pin it in the lane tests.
Allow any positive worker count in the qualification guard without relaxing its coverage-file checks.
Record Bun 1.4.3 measurements and remaining upstream defects in the testing reference and CI comments. Keep integration unisolated while the runner-hang issue remains open; unit-lane soak does not change that gate.
Test Plan
bun run check passes
bun run test passes
bun run build passes
New UI strings use @mangostudio/shared/i18n (no hardcoded strings)
Reported local validation: bun run check; root scripts, 3,517 passed and 13 skipped. Restoring --parallel=1 makes both updated lane assertions fail on the expected worker flag.
Current-head CI passed tooling/typecheck, all four coverage shards, frontend tests, build and automated smoke checks at fa1d578. Full local bun run test, bun run build and manual smoke were not run. No UI strings were added.
Recorded local measurements on 1.4.3+c6da4a4d3, WSL2 pinned to four cores, 437 unit files, output redirected to a file: coverage shards took 65–69s at one worker and 19–20s at four. All four shards passed at each width with identical file partitions. The unsharded unit soak was 12/12 clean, with 5,380 passed per run.
Both widths covered 61,807 lines; merged coverable lines differed by 51, moving line coverage from 80.15% to 80.20%, within the QA verdict's 0.1pp epsilon. These are local measurements, not a hosted speedup claim. Root-script contention and serial integration limit the expected CI saving.
Twelve clean local runs do not establish hosted reliability at the previous 2/36 failure rate. Watch main-branch shard runs for aborts or watchdog kills and revert the flag if they recur.
Run `test:coverage:unit` at `--parallel=4` instead of one isolated
worker. Bun 1.4.3 carries oven-sh/bun#44581, which closedoven-sh/bun#37968: the `EEXIST: epoll_ctl` abort that lost whole files
under in-process workers. The per-file stdio dup leak behind it is
still there (oven-sh/bun#38008 is unmerged); what the release removed
is the wrong-epoll release that turned the leak into a failure.
Pinned to four cores on 1.4.3, the unit lane at four workers was 12 of
12 clean where it failed 4 of 10 before, and the coverage lane with its
CI flags dropped from 65-69s to 19-20s per shard at four shards. Both
widths split the 437 files identically and cover the same 61,807 lines.
Coverable lines differ by 51 in the merged report, which moves line
coverage 0.05pp, inside the QA verdict's 0.1pp epsilon.
The lane stays one invocation, so it still writes one LCOV. The plain
`test:unit` worker processes, the unisolated integration lane and the
shard watchdog are unchanged: oven-sh/bun#39709 is still open.
Record the measurements and the upstream state in the testing
reference, and correct the comments that still described the fix as
pending.
Signed-off-by: Julio Polycarpo <julio@polycarpo.dev>
juliopolycarpo
added
type: test
Test coverage, test infrastructure, or test reliability work.
type: ci
CI, release, checks, automation, or GitHub workflow changes.
area: docs
Documentation tree, guides, reference docs, or contributor docs.
type: perf
Performance-focused changes (matches the perf commit type).
labels
Oct 12, 2026
CI duration (wall clock, critical path, slowest jobs)
CI Duration
Report-only timing from the Actions jobs API; runner variance does not affect the gate.
Comparisons cover the jobs both runs share, so main-only stages never skew the baseline.
Metric
Base
Head
Δ
Wall clock (shared jobs)
15m 31s
12m 6s
🟢 ▼ -3m 25s
Critical path
Distribution / Runtime / Runtime windows-arm64 · 11m 46s
Distribution / Runtime / Runtime windows-x64 · 8m 26s
API/shared branches and statements are source-derived from LCOV line hits because Bun LCOV does not emit branch or statement records.
Rust crates: line, function and region coverage from cargo llvm-cov over the ubuntu libtest run. Branch coverage is not collected and doctests are not instrumented. A crate shows n/a when the Rust lane did not run for the change or its profile data is incomplete.
Component
Metric
Base
Head
Δ
apps/api
lines
89.99% (80,106/89,019)
90.04% (80,106/88,966)
🟢 ▲ +0.05pp
apps/api
statements
80.85% (32,600/40,322)
80.85% (32,600/40,322)
⚪ ▲ = 0
apps/api
functions
87.78% (8,086/9,212)
87.70% (8,081/9,214)
🔴 ▼ -0.08pp
apps/api
branches
55.84% (12,418/22,237)
55.90% (12,430/22,237)
🟢 ▲ +0.06pp
apps/frontend
lines
86.41% (55,878/64,666)
86.41% (55,878/64,666)
⚪ ▲ = 0
apps/frontend
statements
83.49% (15,383/18,424)
83.49% (15,383/18,424)
⚪ ▲ = 0
apps/frontend
functions
80.04% (4,067/5,081)
80.04% (4,067/5,081)
⚪ ▲ = 0
apps/frontend
branches
55.18% (7,646/13,857)
55.18% (7,646/13,857)
⚪ ▲ = 0
apps/shared
lines
89.74% (22,014/24,532)
89.74% (22,014/24,532)
⚪ ▲ = 0
apps/shared
statements
80.25% (4,656/5,802)
80.25% (4,656/5,802)
⚪ ▲ = 0
apps/shared
functions
79.96% (850/1,063)
79.96% (850/1,063)
⚪ ▲ = 0
apps/shared
branches
50.52% (1,564/3,096)
50.52% (1,564/3,096)
⚪ ▲ = 0
crates/mango-protocol
lines
94.82% (7,469/7,877)
n/a
n/a
crates/mango-protocol
functions
94.34% (984/1,043)
n/a
n/a
crates/mango-protocol
regions
94.88% (11,159/11,761)
n/a
n/a
crates/mangostudio-launcher
lines
35.78% (78/218)
n/a
n/a
crates/mangostudio-launcher
functions
36.36% (12/33)
n/a
n/a
crates/mangostudio-launcher
regions
34.94% (123/352)
n/a
n/a
crates/mangostudio-runtime
lines
92.34% (55,410/60,007)
n/a
n/a
crates/mangostudio-runtime
functions
91.05% (6,184/6,792)
n/a
n/a
crates/mangostudio-runtime
regions
92.07% (85,959/93,365)
n/a
n/a
crates/mangostudio-runtime-contract
lines
99.61% (763/766)
n/a
n/a
crates/mangostudio-runtime-contract
functions
100.00% (111/111)
n/a
n/a
crates/mangostudio-runtime-contract
regions
99.64% (1,111/1,115)
n/a
n/a
Lines of Code
Authored source: production and test files of every discovered component. Generated, fixture, config and docs lines are recorded in the envelope but not counted here.
Component
Base
Head
Δ
apps/api
1,313 files / 203,252 lines
1,313 files / 203,252 lines
files ⚪ ▲ = 0 • code ⚪ ▲ = 0
apps/frontend
995 files / 105,245 lines
995 files / 105,245 lines
files ⚪ ▲ = 0 • code ⚪ ▲ = 0
apps/shared
304 files / 40,533 lines
304 files / 40,533 lines
files ⚪ ▲ = 0 • code ⚪ ▲ = 0
crates/mango-protocol
74 files / 15,735 lines
74 files / 15,735 lines
files ⚪ ▲ = 0 • code ⚪ ▲ = 0
crates/mangostudio-launcher
1 files / 284 lines
1 files / 284 lines
files ⚪ ▲ = 0 • code ⚪ ▲ = 0
crates/mangostudio-runtime
248 files / 105,094 lines
248 files / 105,094 lines
files ⚪ ▲ = 0 • code ⚪ ▲ = 0
crates/mangostudio-runtime-contract
8 files / 1,468 lines
8 files / 1,468 lines
files ⚪ ▲ = 0 • code ⚪ ▲ = 0
packages/cli
1 files / 76 lines
1 files / 76 lines
files ⚪ ▲ = 0 • code ⚪ ▲ = 0
packages/protocol
71 files / 9,140 lines
71 files / 9,140 lines
files ⚪ ▲ = 0 • code ⚪ ▲ = 0
scripts
464 files / 67,847 lines
464 files / 67,847 lines
files ⚪ ▲ = 0 • code ⚪ ▲ = 0
total
3,479 files / 548,674 lines
3,479 files / 548,674 lines
files ⚪ ▲ = 0 • code ⚪ ▲ = 0
Frontend Bundle
Metric
Base
Head
Δ
gzip total
1.32 MiB
1.32 MiB
⚪ ▲ = 0
gzip JavaScript
1.30 MiB
1.30 MiB
⚪ ▲ = 0
gzip CSS
20.3 KiB
20.3 KiB
⚪ ▲ = 0
gzip HTML
809 B
809 B
⚪ ▲ = 0
tracked files
121
121
⚪ ▲ = 0
Dependencies
Metric
Base
Head
Δ
locked packages
513
513
⚪ ▲ = 0
direct dependencies
42
42
⚪ ▲ = 0
direct devDependencies
39
39
⚪ ▲ = 0
workspace manifests
5
5
⚪ ▲ = 0
Tests
Single full-suite pass (unit + integration, from the coverage run).
One row per lane, deduplicated by test. ≥ marks a lower bound: a shard report or process receipt was missing, cut off or timed out. A partial or unavailable lane is never counted as zero, and a total over one is never complete.
74f557a perf(test): run the API unit coverage lane at four workers
fa1d578 docs(test): keep the integration lane's isolate gate on bun#39709
Full commit messages
74f557a perf(test): run the API unit coverage lane at four workers
perf(test): run the API unit coverage lane at four workers
Run `test:coverage:unit` at `--parallel=4` instead of one isolated
worker. Bun 1.4.3 carries oven-sh/bun#44581, which closed
oven-sh/bun#37968: the `EEXIST: epoll_ctl` abort that lost whole files
under in-process workers. The per-file stdio dup leak behind it is
still there (oven-sh/bun#38008 is unmerged); what the release removed
is the wrong-epoll release that turned the leak into a failure.
Pinned to four cores on 1.4.3, the unit lane at four workers was 12 of
12 clean where it failed 4 of 10 before, and the coverage lane with its
CI flags dropped from 65-69s to 19-20s per shard at four shards. Both
widths split the 437 files identically and cover the same 61,807 lines.
Coverable lines differ by 51 in the merged report, which moves line
coverage 0.05pp, inside the QA verdict's 0.1pp epsilon.
The lane stays one invocation, so it still writes one LCOV. The plain
`test:unit` worker processes, the unisolated integration lane and the
shard watchdog are unchanged: oven-sh/bun#39709 is still open.
Record the measurements and the upstream state in the testing
reference, and correct the comments that still described the fix as
pending.
Signed-off-by: Julio Polycarpo <julio@polycarpo.dev>
fa1d578 docs(test): keep the integration lane's isolate gate on bun#39709
docs(test): keep the integration lane's isolate gate on bun#39709
The api lane comment tied leaving integration unisolated to the unit
lane soaking at four workers. Those are unrelated: the unit lane's
width rests on oven-sh/bun#44581, while integration stays out of
isolate mode because that machinery is what hangs the runner
(oven-sh/bun#39709, still open) and its files are spawn-heavy.
Restore the present tense and name the open issue as the gate.
Signed-off-by: Julio Polycarpo <julio@polycarpo.dev>
📝 Changelog Preview
Entries this branch would add to the changelog on release:
[Unreleased]
⚡ Performance
(test) Run the API unit coverage lane at four workers
📚 Documentation
(test) Keep the integration lane's isolate gate on bun#39709
The reason will be displayed to describe this comment to others. Learn more.
Stale comment
Verdict: no merge blockers. One-commit PR (74f557a5) on perf/unit-coverage-parallel-workers → main. No prior reviews or threads.
The load-bearing change is apps/apitest:coverage:unit--parallel=1 → --parallel=4. Plain test:unit stays multi-process --parallel=1, integration stays unisolated, shard count and watchdog stay. Lane tests pin --parallel=4; the qualification guard now accepts any positive --parallel count, which is the right split (width is not a coverage-set change). Docs and comments match Bun 1.4.3 / #44581 vs still-open #38008 leak and #39709 hang. This CI check suite is green (51 passed, 0 failed).
P0 / P1: none.
P2: one comment in scripts/lib/test-lanes.ts now ties keeping integration unisolated to “unit lane soak at four workers.” That soak is this PR’s experiment, not the unblock for isolate on spawn-heavy integration (#39709 / #38008). Restore the original gate so a green coverage soak does not read as permission to isolate that lane.
Watch the first few main-branch shard runs for epoll_ctl or watchdog kills; revert the flag if either returns. Not blocking.
Sent by Cursor Automation: Mango Ecosystem CodeReview
The api lane comment tied leaving integration unisolated to the unit
lane soaking at four workers. Those are unrelated: the unit lane's
width rests on oven-sh/bun#44581, while integration stays out of
isolate mode because that machinery is what hangs the runner
(oven-sh/bun#39709, still open) and its files are spawn-heavy.
Restore the present tense and name the open issue as the gate.
Signed-off-by: Julio Polycarpo <julio@polycarpo.dev>
The reason will be displayed to describe this comment to others. Learn more.
Verdict: no merge blockers. Head fa1d578 on perf/unit-coverage-parallel-workers → main (a39ca7e). CI check suite green (51 passed, 0 failed, 20 skipped).
Since last review:fa1d578 (docs) addresses the P2 on scripts/lib/test-lanes.ts. Integration stays unisolated while oven-sh/bun#39709 is open; the comment no longer keys that on unit-lane soak. Present tense restored. Thread marked fully addressed.
The load-bearing change is still apps/apitest:coverage:unit--parallel=1 → --parallel=4. Plain test:unit stays multi-process --parallel=1, integration stays unisolated, shard count and watchdog stay. Lane tests pin --parallel=4; the qualification guard accepts any positive --parallel count (width is not a coverage-set change). Docs match Bun 1.4.3 / #44581 vs still-open #38008 leak and #39709 hang.
P0 / P1 / P2: none remaining.
Watch the first few main-branch shard runs for epoll_ctl or watchdog kills; revert the flag if either returns. Not blocking.
Sent by Cursor Automation: Mango Ecosystem CodeReview
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
area: apiElysia routes, API modules, middleware, or backend services.area: buildBinary build, runtime paths, packaging, or standalone output.area: docsDocumentation tree, guides, reference docs, or contributor docs.type: ciCI, release, checks, automation, or GitHub workflow changes.type: perfPerformance-focused changes (matches the perf commit type).type: testTest coverage, test infrastructure, or test reliability work.
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
juliopolycarpo
added
type: test
github-actions
Bot
added
area: api
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
Summary
Run API unit coverage at four isolated workers per CI shard now that Bun 1.4.3 has removed the known
epoll_ctlabort path. Keep onebun testinvocation and one LCOV per unit shard.Changes
apps/api'stest:coverage:unitto--parallel=4and pin it in the lane tests.Test Plan
bun run checkpassesbun run testpassesbun run buildpassesNew UI strings use
@mangostudio/shared/i18n(no hardcoded strings)Manually smoke-tested: login, chat creation, image generation
Reported local validation:
bun run check; root scripts, 3,517 passed and 13 skipped. Restoring--parallel=1makes both updated lane assertions fail on the expected worker flag.Current-head CI passed tooling/typecheck, all four coverage shards, frontend tests, build and automated smoke checks at
fa1d578. Full localbun run test,bun run buildand manual smoke were not run. No UI strings were added.Screenshots / GIFs
Not applicable.
Notes
1.4.3+c6da4a4d3, WSL2 pinned to four cores, 437 unit files, output redirected to a file: coverage shards took 65–69s at one worker and 19–20s at four. All four shards passed at each width with identical file partitions. The unsharded unit soak was 12/12 clean, with 5,380 passed per run.