github artifact-keeper/artifact-keeper v1.12.3
Artifact Keeper 1.12.3

one hour ago

Artifact Keeper 1.12.3

1.12.3 is a patch release on 1.12: about 15 fixes, two security fixes, two light migrations, no API removals. Upgrade from 1.12.2 directly; from 1.12.1 or earlier, read the 1.12.2 upgrade notes first. The headline fixes are migration jobs that no longer stay running forever after a pod restart (#6253), a hung scanner that no longer holds its scan for half an hour (#5942), the builds API returning what it stores without exposing another user's secrets (#6046, #6281), a Virtual OCI repository that lists its members' signatures (#5792), and wasmtime 36.0.21 for three RUSTSEC advisories (#6273).

Upgrade notes

Migrations

Upgrading from 1.12.2 runs two migrations, both light.

  • 399 (#6090) narrows the conda shard dirty-mark triggers with CREATE OR REPLACE TRIGGER: writes to artifacts wait for a moment; reads and downloads do not. Each attempt waits at most 5 s for its lock, with 5 attempts; under a sustained write load startup fails and can simply be retried.
  • 400 (#6253) adds the claim columns claimed_by, claim_token and claim_expires_at to migration_jobs, and gives every job that is running at that moment an expired claim. A janitor on every 1.12.3 replica pauses running jobs whose claim has lapsed, so migration jobs still running on 1.12.2 pods are paused during the upgrade (error summary worker lost). A 1.12.2 worker whose job is paused stops at its next item. Resume them with /resume once no 1.12.2 pod is left; the resumed job skips the items it already finished. Don't start or resume migration jobs until no 1.12.2 pod is left: a job started on a 1.12.2 pod during the rollout gets no claim, so it is not paused automatically if that pod goes away; /pause it, then /resume it on a 1.12.3 pod.

During a rolling deploy, 1.12.2 pods still running may briefly serve a stale conda owned-name set or channeldata.json after 399 commits. Keep the rollout short.

Rolling back. A 1.12.2 image refuses to start on the migrated database (Migration(VersionMissing(…))). Roll back with SKIP_MIGRATIONS=true; 399 and 400 are compatible with 1.12.2 code (1.12.2 ignores the claim columns, and has no janitor, so a job orphaned by a pod restart stays running again). To also restore 1.12.2's conda guard behaviour, restore migration 298's trigger block first (see the header of migration 399).

New and changed settings

  • SCANNER_PROCESS_TIMEOUT_SECS (new, unset: no limit) caps each local grype/trivy process (#5942). On pods that scan without the queue, set it to STUCK_SCAN_THRESHOLD_SECS (1800 by default) or lower: the janitor fails a hung scan but never stops the process.
  • TRIVY_IGNORE_EMBEDDED_POM_DECLARATIONS (new, default false, behaviour unchanged) stops the Trivy filesystem scan grading dependencies only declared in a JAR's embedded build POM (#6233). Known gap when enabled: a shaded library whose own META-INF/maven wasn't carried into the jar is then not flagged by either engine.
  • Scanner adapter 1.7.1 (#5913) is a separate image. Connected installs download trivy's Java DB (~1 GiB) in the background at startup; leave room for two copies in SCANNER_TRIVY_CACHE_DIR. If it cannot be downloaded, images with JARs fail to scan (retried after 5 minutes) rather than being reported clean.

Behaviour changes operators will notice

  • Listener clashes are resolved at startup (#6213). With METRICS_PORT=9090 and GRPC_PORT unset, /metrics now serves and gRPC is turned off with an ERROR; set GRPC_PORT to another port to keep it. Two explicitly configured listeners on one address, or a METRICS_PORT / explicit GRPC_PORT that cannot bind, now fail startup.
  • A queued scan whose scanner hangs is retried (#5942). Between attempts its row reads failed, so a block_on_fail policy quarantines the artifact and releases it when a retry succeeds.
  • A migration job whose worker is gone is paused, not left running (#6253). A clean shutdown pauses the job; after an OOM kill or eviction the janitor pauses it once its claim lapses (up to about 5 minutes), with the error summary worker lost. Jobs are never resumed automatically; /resume continues them.
  • Unknown /api/v1 paths answer a JSON 404 with the message Not found (#6299); previously the message said Repository not found. The status code is unchanged. The error that stops the backend at startup is now a log line (fatal error, exiting, the error in error), JSON under LOG_FORMAT=json, instead of a plain Error: ... on stderr.
  • Maven and Gradle files outside the groupId/artifactId/version/file layout no longer register a package (#6271). Files such as _remote.repositories, resolver-status.properties or com/myDomain/notes.txt keep their artifact row but no longer show up as a package (com at version myDomain). Rows registered before the upgrade stay; see the cleanup below.
  • conda: a non-package file in a hosted channel no longer counts for the virtual dependency-confusion guard, current_repodata.json or channeldata.json (#6090).
  • npm attestation lookups on hosted repositories answer 404 instead of 501 (#5870).
  • A soft-deleted Maven artifact answers 404 on the legacy flat-key path (#5773).
  • The builds API returns VCS fields and metadata to the build's creator and admins (#6046, #6240). sort_by is now honoured, and credentials in vcs_url are stripped on create. Builds created before 1.12.3 keep any credentials embedded in their stored vcs_url (they are never returned); rotate any CI token that was embedded in a clone URL.
  • On a build someone else created, non-owners see only artifacts that are stored in a repository they can read, and artifact_count matches what they see (#6281). The creator and admins still see every artifact and the full count.
  • A Virtual OCI repository lists its members' referrers (#5792), so cosign verify and oras discover now find signatures stored in a hosted member.
  • Nexus migrations honour date_from/date_to and dry runs count existing artifacts as skipped (#6195, #6196).
  • API token bursts no longer get 503 while the token cache is cold (#6215).

Cleaning up phantom Maven packages (optional)

The backend names Maven and Gradle packages groupId:artifactId, but a name without a colon is not proof of a phantom: older rows named by the bare artifactId are real packages with their files behind them. Treat a row as a leftover only when no artifact in its repository sits at .../<name>/<version>/... for any of its versions. Run the read-only query first and review its output:

SELECT r.key, p.name, p.version FROM packages p JOIN repositories r ON r.id = p.repository_id WHERE r.format::text IN ('maven', 'gradle') AND p.name NOT LIKE '%:%' AND NOT EXISTS (SELECT 1 FROM artifacts a WHERE a.repository_id = p.repository_id AND (a.path LIKE '%/' || p.name || '/' || p.version || '/%' OR EXISTS (SELECT 1 FROM package_versions pv WHERE pv.package_id = p.id AND a.path LIKE '%/' || p.name || '/' || pv.version || '/%'))) ORDER BY r.key, p.name;

If every row listed is one you want gone, delete exactly those rows (their package_versions rows go with them):

DELETE FROM packages p USING repositories r WHERE r.id = p.repository_id AND r.format::text IN ('maven', 'gradle') AND p.name NOT LIKE '%:%' AND NOT EXISTS (SELECT 1 FROM artifacts a WHERE a.repository_id = p.repository_id AND (a.path LIKE '%/' || p.name || '/' || p.version || '/%' OR EXISTS (SELECT 1 FROM package_versions pv WHERE pv.package_id = p.id AND a.path LIKE '%/' || p.name || '/' || pv.version || '/%')));

Both queries scan the repository's artifacts for each candidate row, so they can take a while on a large Maven repository.

Security

  • wasmtime 36.0.21 for RUSTSEC-2026-0566, 0569 and 0571 (#6273). RUSTSEC-2026-0567 and 0573 are not reachable with Artifact Keeper's plugin runtime configuration and are ignored with that reasoning; the wasmtime major upgrade that fixes them is tracked in #6276.
  • The builds API no longer shows a non-owner an artifact because of a path the build's creator typed (#6281), and no longer reveals how many artifacts are hidden from them. VCS fields and metadata go only to the creator and admins, and vcs_url never carries credentials (#6046, #6240).
  • scanner-adapter 1.7.1 bundles trivy rebuilt on Go 1.27.4 for CVE-2026-97054 (#6320).

Formats, proxy and storage

  • Maven: a soft-deleted artifact and its sidecars answer 404 on the legacy flat-key path (#5773); files outside the GAV layout register no package (#6271).
  • OCI: the Referrers API on a Virtual merges its readable members' referrers (#5792).
  • npm: hosted attestation lookups answer 404 (#5870).
  • conda: writes in other formats no longer fire the conda shard triggers, and non-package files no longer affect conda documents (#6090).
  • Proxy: the "cached UNSCANNED" warning fires only for objects with no scan verdict, once per object (#5630).

Scanning, migration and operations

  • A hung scanner is stopped at the queued job's budget and the job retried; SCANNER_PROCESS_TIMEOUT_SECS caps local scanner processes (#5942).
  • Concurrent first scans of images with JARs no longer race on the Java DB download (scanner adapter 1.7.1, #5913); the Trivy filesystem scan can ignore embedded POM declarations (#6233).
  • Migration jobs are paused, not orphaned, when their pod goes away (#6253); Nexus migrations honour the date window (#6195); dry runs count existing artifacts as skipped (#6196).
  • Listeners are resolved and bound before anything is served (#6213); API token validation is single-flight per token (#6215); the fatal startup error is a structured log line (#6299).
  • Development and CI: steadier CI and release pipeline on hosted runners (#6242).

Sponsors

Thank you to our sponsors for supporting ongoing development of Artifact Keeper.

Backers

Become a sponsor to support the project and get your name listed here.

Thank You

Thanks to the external contributors and issue reporters who shaped this release.

Contributed code

  • @nicola-preda: migration jobs paused instead of stuck when their worker is gone (#6295), Maven files outside the GAV layout no longer registering packages (#6286), and the fatal startup error as a JSON log line with Not found on unknown /api/v1 paths (#6301).
  • @macmoonshine: soft-deleted Maven artifacts refused on the legacy flat-key path (#6149) and the scan-on-proxy gap warning scoped to unscanned objects (#5829).
  • @mwindmark: npm attestation lookups on hosted repositories answering 404 (#5927).

Reported the issues this release closes

  • @cervotoc: a deleted Maven artifact still served (#5773), the empty OCI Referrers index on a Virtual (#5792) and npm attestation lookups (#5870).
  • @kalavt: Nexus migrations ignoring the date window (#6195), dry runs counting existing artifacts as completed (#6196) and migration jobs stuck running after a pod restart (#6253).
  • @nicola-preda: the builds API dropping VCS fields, metadata and artifacts (#6046), and the fatal startup error under LOG_FORMAT=json (#6299).
  • @romainwager: a listener that cannot bind leaving the server running without it (#6213) and 503 on API token bursts (#6215).
  • @lucPfa: phantom Maven packages from files outside the GAV layout (#6271, #6306).

Full changelog: v1.12.2...v1.12.3

Don't miss a new artifact-keeper release

NewReleases is sending notifications on new releases.