fix(ci): bound the last unguarded bun installs against a hang

Same defect as #4204/#4206: an erroring bun install (corrupt
tarball extraction) can hang instead of exiting, eating a full
job timeout.

- install-workspace-dependencies.sh: portable bash-only watchdog
  (no GNU `timeout` on macOS/Windows runners) so its own retry
  loop can actually fire on a hang.
- k8s-jobs/Dockerfile.example: same timeout+retry as the other
  two fixed Dockerfiles (Debian base, GNU timeout available).
- ci.yml: route the bare Windows install step through the now
  guarded script instead of a raw bun install.

Nothing else in this class is left.

Claude-Session: https://claude.ai/code/session_018Ne5CfTmFV6GStSKZDf2dy
This commit is contained in:
Miguel Angel Simon Sierra
2026-09-19 19:17:13 -04:00
parent 2d84937e1f
commit 33145fa4b6
3 changed files with 28 additions and 3 deletions
+24 -1
View File
@@ -4,9 +4,32 @@ set -u
max_attempts=3
retry_delay_seconds="${HF_BUN_INSTALL_RETRY_DELAY_SECONDS:-5}"
install_timeout_seconds="${HF_BUN_INSTALL_TIMEOUT_SECONDS:-300}"
# Portable: this runs on ubuntu/macos/windows(Git Bash), and GNU `timeout`
# is not reliably present there (macOS has none by default). A hung bun
# install errors then never exits (see fix/regression-docker-build-timeout),
# so without this the retry loop below never fires.
run_with_timeout() {
local seconds=$1
shift
"$@" &
local pid=$!
local elapsed=0
while kill -0 "$pid" 2>/dev/null; do
if ((elapsed >= seconds)); then
kill -TERM "$pid" 2>/dev/null
wait "$pid" 2>/dev/null
return 124
fi
sleep 1
elapsed=$((elapsed + 1))
done
wait "$pid"
}
for ((attempt = 1; attempt <= max_attempts; attempt += 1)); do
if bun install --frozen-lockfile "$@"; then
if run_with_timeout "$install_timeout_seconds" bun install --frozen-lockfile "$@"; then
exit 0
else
install_status=$?