Skip to content

ci: one heavy test suite per runner VM#2044

Open
cpunion wants to merge 2 commits into
xgo-dev:mainfrom
cpunion:ci-split-heavy
Open

ci: one heavy test suite per runner VM#2044
cpunion wants to merge 2 commits into
xgo-dev:mainfrom
cpunion:ci-split-heavy

Conversation

@cpunion

@cpunion cpunion commented Jul 8, 2026

Copy link
Copy Markdown
Collaborator

Structural fix for the post-backbone runner deaths (ubuntu exit-143 "shutdown signal" ~8 minutes into the heavy-package phase, reproduced on main and every stage-5 branch) and the mac go-list WaitDelay flakes: the covered run stacks four heavy llgo-building test packages (cl, ssa, internal/cabi, test/go) on one VM, and warm shared caches let their compile phases fully overlap.

Design (asymmetric, mindful of scarce mac runners)

  • ubuntu (plentiful runners, where the VMs die): one suite per VM — cl / ssa+cabi / test tree / rest. Peak cannot stack by construction.
  • macos (scarce runners; symptom was slowness, never VM death): two balanced groups of at most two heavy packages each — cl+cabi and ssa+testgo+rest. That restores the pre-backbone peak scale, which mac handled, at the cost of only +1 mac job.

Rejected alternatives: raising budgets (hides it), go test -p 2 (lowers probability only).

Wall time drops to the slowest group; coverage uploads from every job and codecov merges per commit. The embedded-emulator step rides with the cl suites, the std-symbol check with the rest suites.

Branch protection note: required check names change from test (os, 19) to test (os, 19, <suite>).

🤖 Generated with Claude Code

Since the stage-5 backbone merged, the covered test run stacks four
heavy, llgo-building test packages (cl, ssa, internal/cabi, test/go)
on one runner; with the shared go-build cache their compile phases
fully overlap and the concurrent clang/lld burst has been killing
16GB ubuntu runners mid-run ('runner received a shutdown signal',
reproduced across main and every stage-5 branch) and starving mac
runners into go-list WaitDelay expirations (printval/returnorder/
abitype flakes).

Split the test job by suite (cl / ssa+cabi / test tree / rest): each
VM runs one heavy package, so the peak cannot stack by construction,
and wall time drops to the slowest single suite. Coverage still
uploads from every job; codecov merges per commit. The embedded
emulator step rides with cl, the std-symbol check with rest.
@codecov

codecov Bot commented Jul 8, 2026

Copy link
Copy Markdown

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant