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
Commit 432e87e
Browse filesBrowse the repository at this point in the historyBrowse files
ci: gate the expensive jobs on lint, basic and unit (#22234)
* ci: gate the expensive jobs on lint, basic and unit
A pull request that is not going to merge still spent the full CI budget on
it. The seven heavy suites in test.yml were held behind `basic` alone, so a
red `lint` or `unit` launched all 27 integration legs; the eight benchmark
shards were held behind nothing at all, and on #22227 they ran for another 67
minutes after `basic` had already gone red — around 3.3 runner-hours of it.
Hold the test.yml suites behind all three cheap jobs, which are the three a
merge needs anyway. benchmarks.yml becomes callable and test.yml calls it
behind the same three, so a pull request reaches the shards through `needs`
rather than through a job that polls: nothing is spent while the cheap suites
run, and a red one skips the shards outright.
A push to main keeps benchmarks.yml's own trigger and is never gated, because
CodSpeed draws every later comparison against that commit's baseline. Its
concurrency group is named rather than derived from `github.workflow`, which
reads as the caller's name once test.yml calls it.
* ci: cancel superseded pull request runs
Four workflows had no concurrency group, so a new commit left the previous
commit's run going: CodeQL analyses both languages, pkg.pr.new publishes a
preview of code that has been replaced, and dependency-review and the labeler
each run again alongside the run they supersede.
Cancel them on a pull request and leave the `main` runs alone, as code-size
and deep-imports already do — code scanning reads a pull request's alerts
against the default branch's last analysis, and every push to main publishes
its own preview, so neither is superseded by anything.
The labeler is keyed on the pull request number rather than on `github.ref`,
which under `pull_request_target` is the base branch — the usual group would
have every open pull request cancelling the others' labelling.
pr-quality only fires on opened/reopened and release.yml only on a push to a
tag, so a new commit supersedes nothing there.
* test(css): compare the color corpus, and group the parser jobs by language
css-parsing-tests has 8366 cases and webpack compared 13 of them — the whole
color half was run only to confirm no crash, with upstream's expected values
discarded. They are CSSOM serializations rather than minifier output, so they
cannot be compared literally; reading both the input and that serialization
through webpack's own folding and canonicalizing to `#rrggbb` does compare,
wherever both land in sRGB. That is 796 cases, and webpack agrees on all of
them — no exception list.
Where it declines to fold (lab, lch, out-of-gamut) the two spellings are not
claims about the same thing, so those stay out rather than becoming noise:
comparing every case reports 5549 differences, essentially all of them that.
The three jobs reading an upstream parser corpus become one matrix, named for
the language rather than for whose repository the fixtures live in — which had
already drifted, since `html5lib` reads WPT for the corpus html5lib-tests
dropped. They differed only in submodule, suite and coverage flag.
* docs: name syntax-equivalence (chrome) among the coverage uploaders
Its matrix row sets `coverage: true` and the job's Codecov step runs on that
flag, so the list of suites to wait for was one short — and the sentence after
it said everything unlisted cannot move the number. A contributor following it
could read a coverage result as final while that job was still running.
0 commit comments