Repository navigation
Description
Summary
Since #126667 moved the Jest shard setup into a parallel: group, shards sometimes fail before jest runs with:
SyntaxError: Unexpected end of JSON input
at JSON.parse (<anonymous>)
at new Context (/home/runner/work/_actions/actions/download-artifact/d3f86a106a0bac45b974a628896c90dbdf5c8093/dist/index.js:4845:37)
This is a race in the GitHub Actions runner: every step start truncates and rewrites the shared $GITHUB_EVENT_PATH file. Any action that loads @actions/github parses that file at startup, so it crashes if a sibling step in the group starts at the same moment. The bug is reported upstream as actions/runner#4641; the fix PRs (actions/runner#4643, actions/runner#4675) are still open with no maintainer response.
We can fix it on our side without giving up the parallel: speedup.
Evidence
Failures
I checked every frontend.yml run that ended in failure between 2026-09-24 and 2026-10-15:
| Date (UTC) | Job | Failed step | Action that crashed |
|---|---|---|---|
| 2026-10-09 09:59 | Jest (2, 8) (PR) | Download jest-balance.json | dawidd6/action-download-artifact |
| 2026-10-13 17:23 | Jest (5, 8) (push to master) | Download jest-test-files.json | actions/download-artifact |
| 2026-10-15 03:01 | Jest (2, 8) (push to master) | Download jest-test-files.json | actions/download-artifact |
- The same error appears zero times from Sep 24 to Oct 7, when these steps ran one after another. All three cases are after perf(ci): Fetch jest shard dependencies in parallel #126667 merged (2026-10-07 15:46 UTC).
- Treat 3 as a lower bound. I only scanned runs whose final result was a failure, so runs that someone re-ran green aren't counted.
In each case the action crashes inside new Context(), 0.2–0.9s after it starts, while other steps in the group are still starting. Timeline from job 116223369053:
03:01:04.295 Run actions/cache <- "jest transform cache" starts
03:01:04.317 Run actions/download-artifact <- "Download jest-test-files.json" starts, node boots
03:01:04.318 Download action repository 'actions/setup-node' <- setup-node-pnpm composite is starting
03:01:04.495 SyntaxError: Unexpected end of JSON input <- download-artifact reads an empty event.json
at new Context (.../actions/download-artifact/.../dist/index.js:4845:37)
03:01:04.618 Run dawidd6/action-download-artifact <- "Download jest-balance.json" starts
In job 113768151765 the crash is in dawidd6, through the same code path:
SyntaxError: Unexpected end of JSON input
at JSON.parse (<anonymous>)
at new Context (.../dawidd6/action-download-artifact/ac66b43.../node_modules/@actions/artifact/node_modules/@actions/github/lib/context.js:15:37)
Mechanism (runner v2.337.0, which our jobs run on)
ActionRunner.RunAsynccallsWriteWebhookPayload()for every step start. That includesrun:steps and each inner step of a composite action.WriteWebhookPayloaddoesFile.WriteAllText(<temp>/_github_workflow/event.json, …), which truncates the file and then writes it. The file is shared by all steps in the job.@actions/githubcreatescontext = new Context()when the module loads. The constructor doesJSON.parse(readFileSync(process.env.GITHUB_EVENT_PATH)), so it throws if it reads during the truncation window.
Which steps in our group read the event file
| Step in the group | Reads $GITHUB_EVENT_PATH at startup? |
|---|---|
actions/setup-node (first step of .github/actions/setup-node-pnpm) |
yes: the entry module requires @actions/github at the top level (dist/setup/index.js) |
actions/download-artifact |
yes: dist/index.js#L4845 |
dawidd6/action-download-artifact |
yes: through @actions/github in its node_modules |
pnpm/action-setup, actions/cache (restore and save), bash run: steps, gh CLI |
no |
Post steps aren't affected. The runner waits for background steps to finish and then runs post steps one at a time.
Ruled out: giving each step its own copy of the event file
Setting env: GITHUB_EVENT_PATH: <copy> on a step has no effect. The runner applies the step's env: first (ActionRunner.cs#L251) and then overwrites all GITHUB_* variables from the github context (NodeScriptActionHandler.cs#L42-L52). This is also actions/runner#701.
Proposed solutions
1. Recommended: keep parallel:, but run every event-file reader before it
- uses: actions/checkout@…
# Actions that load @actions/github (setup-node, download-artifact, checkout, github-script, …)
# parse $GITHUB_EVENT_PATH at startup, and the runner truncates and rewrites that file every
# time any step starts (actions/runner#4641). Keep them out of the parallel group below.
- uses: actions/setup-node@1e60f620b9541d16bece96c5465dc8ee9832be0b # v4
with:
node-version-file: .node-version
- parallel:
- uses: ./.github/actions/setup-node-pnpm
with:
setup-node: 'false' # new input, defaults to 'true' so the other 19 callers don't change
- name: jest transform cache
uses: actions/cache@… # unchanged
- name: Download jest artifacts
env:
GH_TOKEN: ${{ github.token }}
GH_REPO: ${{ github.repository }}
run: ./.github/workflows/scripts/download-jest-artifacts.sh- In
.github/actions/setup-node-pnpm/action.yml: add asetup-nodeinput (default'true'), and addif: inputs.setup-node == 'true'to theactions/setup-nodestep. - Replace both download actions with a single
ghstep, which never reads the event file. The two downloads also get retries this way; see "Related" below.
Script sketch (not yet tested):
#!/usr/bin/env bash
# Download jest-test-files.json (uploaded by "Configure Jest" in this run) and
# jest-balance.json (latest successful jest-balance.yml run on master).
set -euo pipefail
retry() {
local attempt
for attempt in 1 2 3; do
"$@" && return 0
echo "::warning::Attempt $attempt failed: $*" >&2
sleep $((attempt * 5))
done
return 1
}
# Extract to a fresh dir so a retry never trips over a partial earlier attempt.
download() { # <run-id> <artifact-name> <dest-dir>
local tmp
tmp=$(mktemp -d)
gh run download "$1" --name "$2" --dir "$tmp" && mkdir -p "$3" && cp -R "$tmp"/. "$3"/
}
retry download "$GITHUB_RUN_ID" jest-test-files .
balance_run=$(retry gh run list --workflow jest-balance.yml --branch master --status success \
--limit 1 --json databaseId --jq '.[0].databaseId // empty')
if [[ -z "$balance_run" ]]; then
# Every shard sees the same result here, so they all fall back to splitting by name.
echo "::warning::No successful jest-balance.yml run on master; shards will split by file name" >&2
exit 0
fi
retry download "$balance_run" jest-balance.json tests/js/test-balancer- Downloading
jest-balance.jsonmust not silently fail. If one shard is missing the file,createJestConfig.tsmakes that shard split tests by filename while the other shards split by duration. Some tests would then run twice and others not at all, with nothing failing. - Mid-run downloads work. v4 artifacts are available through the REST API as soon as the upload finishes, so
gh run download "$GITHUB_RUN_ID"can fetchjest-test-fileswhile the run is still going. - Cost: roughly nothing.
- In a passing shard (job 116004543769), the
setup-node-pnpmbranch took 21s and was the critical path;setup-nodewas about 6s of it. - The other branches took 5s or less.
- Running
setup-nodefirst gives about 6s + 13s, close to today's 21s, so the ~9s saving from perf(ci): Fetch jest shard dependencies in parallel #126667 is kept.
- In a passing shard (job 116004543769), the
2. Smaller change: replace only the two download actions with the gh step
This leaves actions/setup-node as the only reader in the group. It starts 0.4–2.5s after the group begins, inside the composite, and is usually past the point where the other branches start. But in job 115437592082, setup-node and dawidd6 started 24ms apart. So this option depends on timing and doesn't guarantee the race is gone.
3. Revert #126667
This also works, but loses about 9s per shard.
Upstream
Upvote or ping actions/runner#4643 or actions/runner#4675. Once one of them merges, the readers can move back into the group.
Related, but a different cause
The same two download steps fail for other reasons too. Between 2026-10-05 and 2026-10-15:
Unknown error: {}: 18 Jest jobs. The balance zip download gets an HTTP error from GitHub's artifact storage, and@octokit/request8.4.1 inside dawidd6 v11 hides the response body. This started Oct 5, before perf(ci): Fetch jest shard dependencies in parallel #126667.Server Error: 15 Jest jobs, during GitHub's 2026-10-12 Actions incident.
Neither is caused by the race. The retries in option 1 should cover both, and gh will print the real HTTP status when it does fail.
Activity
linear-code commentedon Oct 15, 2026
ENG-8821
fix(ci): Keep event-file readers out of parallel and background steps
fix(ci): Keep event-file readers out of parallel and background steps (…
ci(jest): Run setup-node-pnpm before the parallel setup group (#128035)
fix(ci): Keep event-file readers out of parallel and background steps (…
ci(jest): Run setup-node-pnpm before the parallel setup group (#128035)
ci(acceptance): Start devservices and per-worker Snuba bootstrap early (