Skip to content

Commit fa55d0d

Browse files
committed
ci: move Windows off 24.15.0, which carries the crash we were chasing
The 0xC0000409 aborts are nodejs/node#56645: node calling uv_async_send() on a handle already marked UV_HANDLE_CLOSING, which trips "Assertion failed: !(handle->flags & UV_HANDLE_CLOSING), win/async.c:76" and exits 3221226505 - 0xC0000409, the same code we have been seeing. Everything we observed matches it. Windows only, absent on 22.x and present from Node 23 onwards, which is exactly how the matrix behaved. It is a race, which is why it fired at three unrelated tests at three different elapsed times, why --report-on-fatalerror never caught it, and why the same commit crashed once and passed on re-run. It is provoked by sockets being opened and closed, which is most of what this test suite does. So it is not accumulation and not a node-opcua bug, and there is nothing to report upstream: nodejs/node#61999 fixed it on 2026-07-24 and it was backported to 24.x on 2026-08-03/04. 24.19.0 is the first 24.x release that can contain it. The pin to 24.15.0 was right when written and had since inverted. It avoided a libuv fs-event regression in 24.16.0 (nodejs/node#63638) that is itself fixed now (#65118), at the cost of holding us on the last release still carrying the async-close assertion. Kept as an exact version rather than 24.x, and the comment says why: Windows has produced two version-specific libuv regressions in a row and this job doubles as the watch for that class of bug. A watch that reports "some Node 24" is not much of a watch, and the floating tag will have moved by the time anyone reads the failure. continue-on-error stays for now, scoped to the Node 24 entry, and should be removed once 24.19.0 has shown the crash is gone.
1 parent 5a1ac1e commit fa55d0d

1 file changed

Lines changed: 33 additions & 7 deletions

File tree

‎.github/workflows/workflow.yml‎

Lines changed: 33 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -72,7 +72,7 @@ jobs:
7272
build_on_windows:
7373
runs-on: windows-latest
7474
timeout-minutes: 45
75-
# Node 24.15.0 on Windows dies with a native abort - exit -1073740791
75+
# Kept non-gating for now. Node 24.15.0 on Windows died with a native abort - exit -1073740791
7676
# (0xC0000409, STATUS_STACK_BUFFER_OVERRUN) - part way through the end-to-end
7777
# suite, during test_e2e_985 (server restarts with a rotated private key and
7878
# certificate) right after an asymmetricVerify against the new key. It is not
@@ -84,16 +84,41 @@ jobs:
8484
# is diagnosed it must not block merges - but it keeps running, and keeps
8585
# uploading its node-report artifact, because that report carries the native
8686
# stack needed to find the cause.
87-
continue-on-error: ${{ matrix.node-version == '24.15.0' }}
87+
# Still scoped to the Node 24 entry while 24.19.0 proves itself; if the crash
88+
# does not return, this guard should be removed so the job gates again.
89+
continue-on-error: ${{ matrix.node-version == '24.19.0' }}
8890
strategy:
8991
# Don't cancel the whole matrix when one Node version fails.
9092
fail-fast: false
9193
matrix:
9294
# Node 18/20 dropped: pnpm 11 requires a recent Node (>=22.13 in our setup).
9395
# node-version: [20.x, 22.x, 24.x]
94-
# 24.15.0 pinned: libuv regression in 24.16.0+ crashes on
95-
# Windows fs-event (nodejs/node#63638, libuv/libuv#5010).
96-
node-version: [22.x, 24.15.0]
96+
# Was pinned to 24.15.0 to avoid a libuv fs-event regression in 24.16.0
97+
# (nodejs/node#63638, libuv/libuv#5010). That regression is fixed - node PR
98+
# #65118 - and the pin had since inverted: 24.15.0 is the last 24.x still
99+
# carrying the uv_async_send-after-uv_close assertion (nodejs/node#56645,
100+
# "Assertion failed: !(handle->flags & UV_HANDLE_CLOSING), win/async.c:76"),
101+
# which aborts as exit 3221226505 / 0xC0000409. That is the abnormal exit
102+
# this job has been hitting: Windows only, absent on 22.x, present from 23
103+
# onwards, and racy - so it fires at an unrelated test each time and the same
104+
# commit can crash once and pass on re-run. node-opcua opens and closes
105+
# sockets constantly, which is exactly what provokes it.
106+
#
107+
# Fixed by nodejs/node#61999, merged 2026-07-24 and backported to 24.x on
108+
# 2026-08-03/04, so 24.19.0 is the first 24.x release that can contain it.
109+
#
110+
# Why an exact version here and not 24.x, when the Linux job floats:
111+
# Windows has now produced two version-specific libuv regressions back to
112+
# back - fs-event in 24.16.0 (nodejs/node#63638) and the async-close
113+
# assertion present through 24.15.0 (nodejs/node#56645). Both were Windows
114+
# only and neither reproduced on Linux. This job doubles as the watch for
115+
# that class of bug, and a watch is only useful if it names the version it
116+
# ran: with 24.x a crash tells you "some Node 24", and by the time anyone
117+
# looks the floating tag has moved. Pinning keeps a failure attributable and
118+
# makes a bisect between two known versions possible.
119+
#
120+
# Bump deliberately, and re-check those two issues when you do.
121+
node-version: [22.x, 24.19.0]
97122

98123
steps:
99124
- uses: actions/checkout@v4
@@ -184,8 +209,9 @@ jobs:
184209
fail-fast: false
185210
matrix:
186211
# Node 18/20 dropped: pnpm 11 requires a recent Node (>=22.13 in our setup).
187-
# Deliberately NOT pinned to 24.15.0 like the Windows job: every
188-
# abnormal-exit crash on record came from build_on_windows (24.15.0).
212+
# Deliberately unpinned, unlike the Windows job: every abnormal-exit crash
213+
# on record came from build_on_windows, and the underlying bug is
214+
# Windows-only (nodejs/node#56645).
189215
# Linux has never reproduced it, so pinning here would only cost us
190216
# coverage of current Node 24.
191217
node-version: [22.x, 24.x]

0 commit comments

Comments
 (0)