fix(release): publish the manifest only after assets are proven fetchable

Today's outage window came from ordering, and the ordering was baked into the
publish script itself: it pushed main — the branch nodes read the manifest
from — together with the tag, up front, then uploaded and verified assets
afterward. So the manifest advertised the new version for the entire
upload+verify window. When an upload failed inside that window, every polling
node briefly saw a v1.7.126-alpha update whose binary 500'd and whose tarball
did not yet exist.

Reordered so the manifest goes live last:
  1. push the TAG only (the Gitea release and asset URLs hang off it; the tag
     alone changes nothing for nodes)
  2. upload assets
  3. verify every asset downloads in full and matches the manifest sha256/size
  4. only then push main — the step that actually triggers nodes

Also fixes a way a bad asset could slip through unnoticed: the inline
verification ran in a `while read` pipe subshell, where its `fail` (exit 1)
terminated only the subshell and let the script continue to "published and
verified". Verification now runs in the main shell via a new
check-release-assets.sh, which fails hard on the first bad asset. The same
script is the reusable by-hand verifier used to recover today's release
(both assets confirmed 200 + sha256-match before the manifest was re-published).

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
archipelago
2026-08-07 18:05:55 -04:00
co-authored by Claude Opus 5
parent 35f992fdb4
commit 308f3cbd84
2 changed files with 128 additions and 20 deletions
+105
View File
@@ -0,0 +1,105 @@
#!/usr/bin/env bash
# check-release-assets.sh — prove a release's artifacts are actually fetchable
# BEFORE its manifest goes live on main.
#
# The manifest is the trigger: nodes read releases/manifest.json from branch
# main, and the moment it names a new version they try to download it. So the
# assets must resolve before the manifest lands, not after. On 2026-08-07 the
# order was reversed — the v1.7.126-alpha manifest went live while its binary
# 500'd and its frontend tarball had never uploaded — and every polling node
# would have advertised an update it could not fetch.
#
# For each component in the manifest this checks:
# 1. the download URL returns HTTP 200
# 2. the downloaded bytes match the manifest's sha256 and size
#
# It downloads each asset in full, because a HEAD 200 is not proof the body is
# intact — the corrupt binary that day passed HEAD-shaped checks and still
# served a broken stream. Slower, but this is the last gate before publish.
#
# Usage:
# scripts/check-release-assets.sh # check releases/manifest.json
# scripts/check-release-assets.sh path/to/manifest.json
#
# Exit 0 = every asset is downloadable and matches. Non-zero = do NOT publish.
set -euo pipefail
MANIFEST="${1:-releases/manifest.json}"
if [[ ! -f "$MANIFEST" ]]; then
echo "ERROR: manifest not found: $MANIFEST" >&2
exit 2
fi
command -v python3 >/dev/null 2>&1 || { echo "ERROR: python3 required" >&2; exit 2; }
command -v sha256sum >/dev/null 2>&1 || { echo "ERROR: sha256sum required" >&2; exit 2; }
TMP="$(mktemp -d)"
trap 'rm -rf "$TMP"' EXIT
# Emit "url<TAB>sha256<TAB>size<TAB>name" per component, tolerating the field
# name variations the manifest has used (download_url/url, size_bytes/size).
rows="$(python3 - "$MANIFEST" <<'PY'
import json, sys
d = json.load(open(sys.argv[1]))
comps = d.get("components") or []
if not comps:
sys.exit("manifest has no components")
for c in comps:
url = c.get("download_url") or c.get("url") or ""
sha = c.get("sha256") or ""
size = c.get("size_bytes") or c.get("size") or ""
name = c.get("name") or "(unnamed)"
if not url or not sha:
sys.exit(f"component {name!r} missing url or sha256")
print(f"{url}\t{sha}\t{size}\t{name}")
PY
)"
version="$(python3 -c 'import json,sys; print(json.load(open(sys.argv[1])).get("version","?"))' "$MANIFEST")"
echo "Checking release assets for v${version} ($MANIFEST)"
echo ""
fail=0
n=0
while IFS=$'\t' read -r url sha size name; do
[ -z "$url" ] && continue
n=$((n + 1))
out="$TMP/asset.$n"
echo " [$name]"
echo " $url"
code="$(curl -sL -o "$out" -w '%{http_code}' "$url" || echo "000")"
if [ "$code" != "200" ]; then
echo " FAIL: HTTP $code (asset not served)"
fail=1
continue
fi
got_size="$(stat -c%s "$out")"
if [ -n "$size" ] && [ "$size" != "$got_size" ]; then
echo " FAIL: size $got_size, manifest says $size"
fail=1
continue
fi
got_sha="$(sha256sum "$out" | awk '{print $1}')"
if [ "$got_sha" != "$sha" ]; then
echo " FAIL: sha256 mismatch"
echo " served: $got_sha"
echo " manifest: $sha"
fail=1
continue
fi
echo " OK: HTTP 200, ${got_size} bytes, sha256 matches"
done <<< "$rows"
echo ""
if [ "$fail" -ne 0 ]; then
echo "REFUSING: one or more assets are not fetchable or do not match the manifest."
echo "Do NOT publish the manifest — nodes would advertise an update they cannot"
echo "apply. Upload/repair the assets, re-run this, and only then flip the"
echo "manifest live on main."
exit 1
fi
echo "OK: all $n asset(s) for v${version} download and match the manifest."
+23 -20
View File
@@ -62,8 +62,19 @@ repo_path=${repo_path%.git}
api="$scheme://$host/api/v1/repos/$repo_path"
release_url="$api/releases/tags/v${VERSION}"
echo "Pushing main and v${VERSION} to $REMOTE..."
git -C "$PROJECT_ROOT" push "$REMOTE" main "refs/tags/v${VERSION}"
# ORDER MATTERS. The manifest is the trigger — nodes read releases/manifest.json
# from branch main and try to download the named version the moment it appears.
# So main (which carries the live manifest) must be pushed LAST, only after the
# assets are uploaded and their bytes verified against the manifest. The tag is
# pushed first because the Gitea release and its asset download URLs hang off it,
# but the tag alone changes nothing for nodes.
#
# This used to push main and the tag together, up front, then upload assets. That
# left the manifest live for the entire upload+verify window — and on 2026-08-07
# an upload failed inside that window, so every polling node briefly advertised a
# v1.7.126-alpha update whose binary 500'd and whose tarball did not exist.
echo "Pushing tag v${VERSION} to $REMOTE (not main yet)..."
git -C "$PROJECT_ROOT" push "$REMOTE" "refs/tags/v${VERSION}"
release_json=$(curl -fsS -u "$auth" "$release_url" || true)
if [ -z "$release_json" ]; then
@@ -107,24 +118,16 @@ upload_asset() {
upload_asset "$BACKEND" "archipelago"
upload_asset "$FRONTEND" "archipelago-frontend-${VERSION}.tar.gz"
echo "Verifying public download URLs from manifest (size + sha256)..."
python3 - "$PROJECT_ROOT/releases/manifest.json" <<'PY' | while read -r url size sha; do
import json
import sys
echo "Verifying public download URLs (full GET + size + sha256)..."
# Delegated to check-release-assets.sh so the same verifier is used here and by
# hand during recovery. It fails hard on the first bad asset — the previous
# inline `while read` ran in a pipe subshell, where a `fail` (exit) killed only
# the subshell and let this script march on to "published and verified".
"$PROJECT_ROOT/scripts/check-release-assets.sh" "$PROJECT_ROOT/releases/manifest.json" \
|| fail "asset verification failed — NOT pushing main. The manifest stays off the branch nodes read, so no node sees a version it cannot fetch. Repair the assets and re-run."
manifest = json.load(open(sys.argv[1]))
for component in manifest["components"]:
print(component["download_url"], component["size_bytes"], component["sha256"])
PY
# Full GET, not HEAD: a size-correct/content-wrong mirror asset must fail
# the publish gate, so compare the actual bytes against the manifest sha256.
tmp=$(mktemp)
curl -fsSL --max-time 900 -o "$tmp" "$url" || { rm -f "$tmp"; fail "download URL failed: $url"; }
actual_size=$(stat -c %s "$tmp")
actual_sha=$(sha256sum "$tmp" | awk '{print $1}')
rm -f "$tmp"
[ "$actual_size" = "$size" ] || fail "download size mismatch for $url (expected $size, got $actual_size)"
[ "$actual_sha" = "$sha" ] || fail "download sha256 mismatch for $url (expected $sha, got $actual_sha)"
done
# Assets are proven fetchable — only now does the manifest become live.
echo "Assets verified. Pushing main to $REMOTE (this makes v${VERSION} live)..."
git -C "$PROJECT_ROOT" push "$REMOTE" main
echo "Release v${VERSION} published and verified on $REMOTE."