deploy: gapless additive deploys, and verify the bundle not just the page

The outage: deploy did rm -rf on the web root, THEN docker cp. Anyone loading
in that window — or holding a cached index.html from before it — asked for a
hashed bundle that no longer existed, and nginx answered with HTML: "Failed to
load module script... MIME type of text/html". Exactly what got reported.

Two changes:
- Merge instead of swap. Hashed bundles never collide, so old builds' assets
  stay alive for stale HTML while the new index.html takes over. No gap, no
  orphaned caches. Disk is a few MB per deploy; prune by hand if it ever
  matters.
- The verify step now follows the chain the browser follows: fetch the live
  index.html, extract the script src, and assert it serves as
  application/javascript. A 200 on the page proved nothing — the page was
  200 all through the outage.

Verified live: page → index-CmCk0pc0.js → application/javascript, and the
game boots to the title screen on partly.party/toastsim.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
monster 2026-07-22 11:04:57 +10:00
parent d1645bae6e
commit b13750f88a

28
scripts/deploy.sh Normal file → Executable file
View File

@ -19,12 +19,24 @@ echo "[2/4] Rsync dist → $VPS:$STAGING ..."
ssh "$VPS" "rm -rf '$STAGING' && mkdir -p '$STAGING'"
rsync -az dist/ "$VPS:$STAGING/"
echo "[3/4] Swap into $CONTAINER:$REMOTE_WEB_ROOT ..."
ssh "$VPS" "docker exec '$CONTAINER' rm -rf '$REMOTE_WEB_ROOT' \
&& docker cp '$STAGING' '$CONTAINER:$REMOTE_WEB_ROOT'"
echo "[3/4] Merge into $CONTAINER:$REMOTE_WEB_ROOT ..."
# ADDITIVE on purpose — no rm, no gap. The old flow (rm -rf, then docker cp)
# had a downtime window, and worse: any browser/CDN holding yesterday's
# index.html pointed at hashed bundles the rm had just deleted → the classic
# "module script served as text/html" outage. Hashed assets never collide, so
# merging keeps every old build's bundles alive for stale HTML while the new
# index.html takes over for everyone else. Disk cost is a few MB per deploy;
# prune assets/ older than ~30 days by hand if it ever matters.
ssh "$VPS" "docker exec '$CONTAINER' mkdir -p '$REMOTE_WEB_ROOT' \
&& docker cp '$STAGING/.' '$CONTAINER:$REMOTE_WEB_ROOT/'"
echo "[4/4] Verify..."
code=$(curl -so /dev/null -w '%{http_code}' "https://partly.party/toastsim/?cb=$(date +%s)")
echo "https://partly.party/toastsim/ → HTTP $code"
[ "$code" = 200 ] || { echo "NOT DEPLOYED (non-200)"; exit 1; }
echo "done."
echo "[4/4] Verify the page AND the bundle it references..."
page=$(curl -s "https://partly.party/toastsim/?cb=$(date +%s)")
src=$(printf '%s' "$page" | grep -oE 'src="[^"]*\.js"' | sed 's/src="//;s/"//' | head -1)
[ -n "$src" ] || { echo "NOT DEPLOYED (no script tag in live index.html)"; exit 1; }
ctype=$(curl -so /dev/null -w '%{content_type}' "https://partly.party$src")
echo "https://partly.party/toastsim/ → script $src$ctype"
case "$ctype" in
application/javascript*|text/javascript*) echo "done." ;;
*) echo "NOT DEPLOYED (bundle served as $ctype — stale index or missing asset)"; exit 1 ;;
esac