The substring grep for the archive line also captured the archive's
.sbom.json entry, whose file is deliberately not downloaded in this
lane, so the sha256sum check failed on every run. awk now matches the
manifest filename column exactly.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Findings from an adversarial review (Codex, gpt-5.6-sol at max effort)
of 2ff594f4b and 4c34d2309, each independently verified before fixing:
- SBOM generation: buildx attaches a provenance attestation, so every
per-arch digest names an OCI index; Syft's platform default on an
amd64 runner cannot resolve an arm64-only index and the step dies.
Pass --platform explicitly on all four Syft calls (the two classic
lanes had the same latent defect - the renamed workflow has not run
yet, which is why it never fired).
- Release ordering: the HEALTHCHECK survival check now runs against
the pushed architecture image before the versioned and rolling
multi-arch manifests are created, so a broken health config blocks
their promotion; the comment now states honestly that the
arch-suffixed tags are already public at that point.
- Gate assertions: tar's member-argument mode exits non-zero on any
missing name, which under pipefail masked a found forbidden file
when exactly one of them existed; -tv prints symlinks as
'name -> target', defeating $-anchored greps; and the licenses
check proved only one-of-three. Export the rootfs once and assert
every required and forbidden entry individually (busybox/sh and
usr/bin/mc[li] now covered), and match the image healthcheck as an
exact array instead of a substring.
- Probe target vs CLI-configured servers: a probe process cannot see
PID 1's argv, so --url gains EnvVar MINIO_HEALTHCHECK_URL as the
documented way to point the baked-in HEALTHCHECK at a server whose
address/TLS comes from command-line arguments (verified end to end:
server on --address :9010, env var alone turns the container
healthy). Baseline regenerated for the new env token.
- IPv6 zone identifiers: serialize probe URLs via url.URL.String()
so [fe80::1%eth0]:9000 becomes a valid %25-escaped URL (tests added).
- Boolean flags: read --json/--quiet via Bool() so --json=false is
false, instead of IsSet() which treats any occurrence as true.
- Docker's HEALTHCHECK timeout raised to 10s: an outer deadline equal
to the probe's own 5s always SIGKILLed the probe before it could
print its diagnostic line.
- test-release path filter now also triggers on cmd/healthcheck-main.go
and cmd/main.go, so subcommand regressions run the image gate.
Not adopted: require_text's comment-insensitivity in verify-rebrand.sh
(snapshot-tripwire by design, consistent with its other assertions -
the semantic check lives in the CI gate now), and full
staging-then-promote tag publishing (a workflow-wide redesign shared
with the classic lanes, tracked as follow-up).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Publish pgsty/silo:<RELEASE>-distroless (and a rolling :distroless tag)
alongside the classic image: gcr.io/distroless/static-debian12 plus
exactly one program, the silo binary. No shell, no mc, no curl, no
entrypoint script - the binary is the ENTRYPOINT and the baked-in
exec-form HEALTHCHECK runs 'silo healthcheck ready'. The classic image
and its mc-based health checks are deliberately unchanged.
Design: silo.pgsty.com/compatibility/feature/healthcheck/
/data is created in the image layer, world-writable, because Docker
seeds fresh volumes from the layer mountpoint and no entrypoint exists
to repair ownership at runtime (issue #55); the parent directory is
copied from a throwaway busybox stage since COPY of a directory copies
contents rather than the entry itself, which would silently leave
/data at root:0755 and break every non-root run. The MINIO_USERNAME
drop-user path is not supported in this variant; use --user.
test-release.yml now builds the real Dockerfile.distroless on every
gate run (it has no download stages, so it stays offline) and asserts:
the HEALTHCHECK survives into the image config, /data ships 0777, no
shell and no /usr/bin/minio are present, Docker's health state turns
healthy from the baked probe alone, the probe binary execs without a
shell, and SIGTERM still stops the server gracefully - as root and as
--user 1001:1001. docker-release.yml gains the distroless build lanes,
multi-arch manifests, SBOM and provenance attestations, and a
release-blocking check that the pushed manifest still carries the
HEALTHCHECK (a Docker extension absent from the OCI image spec).
Verified locally on linux/arm64: full gate assertions plus bare
'docker run ... healthcheck' exit-code semantics and --version
passthrough.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The RPM, DEB and APK payloads carried no license material at all, and
the OCI image copied LICENSE and CREDITS but not NOTICE. For an AGPL
distribution that is not defensible packaging. Every delivery channel
now carries the same license set:
- nfpm packages install LICENSE and NOTICE under /usr/share/doc/silo/.
The entries are declared once per packager because nfpm only honors
type: license on rpm and silently drops such entries from deb and apk
-- the rpm keeps its %license flag, deb and apk carry plain files at
the same path. CREDITS stays out of the OS packages deliberately: at
1.8MB it would dominate an otherwise ~40MB payload.
- The OCI image gains /licenses/NOTICE next to the existing LICENSE and
CREDITS, in the Dockerfile and in both context-staging paths (the
make docker target and docker-release.yml).
- The GoReleaser archives list their files explicitly; the default
globs would keep missing NOTICE.
NOTICE's first line now reads 2015-2025 to match the banner constants
fixed in b57275be3 -- it had kept upstream's stale 2015-2023 claim --
and verify-rebrand.sh pins both NOTICE copyright lines.
The payload locks move in the same commit: sign-release-rpms.sh expects
the six-file listing, and the test-release validation asserts presence,
count, the rpm %license flag, and sha256 equality of the packaged
LICENSE and NOTICE against the repository files across all three
formats and both architectures.
Also dropped the unit's WorkingDirectory=/usr/local, a leftover from
upstream's /usr/local/bin install layout. Debian policy 9.1.2 keeps
packages out of /usr/local entirely; the binary already installs to
/usr/bin and the service does not depend on its working directory.
Verified locally with nfpm v2.47.0 (the CI-pinned version): built all
six packages via package-release.sh, then ran the test-release
validation step verbatim (extracted from the workflow YAML) and the
signing script's payload comparison against them in an Ubuntu 24.04
container -- both pass. Built the arm64 image from the real
Dockerfile.goreleaser and confirmed /licenses holds CREDITS, LICENSE
and NOTICE. make rebrand-guard passes with the new NOTICE pins.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Feng Ruohang <rh@vonng.com>
actions/checkout@v4 and actions/setup-go@v5 declare runs.using: node20, so
every Go CI job printed the runner's Node 20 deprecation warning. The
release, test-release and docker-release workflows carried six more node20
actions that never appeared in that warning only because those workflows
run on a tag or a dispatch rather than on every push: upload-artifact v4,
goreleaser-action v6, build-push v6 and the docker qemu/buildx/login v3
line. All of them move to the current majors, which declare node24. The
SHA-pinned rest already qualified -- attest and sbom-action are node24,
cosign-installer is composite.
Checked against what those majors removed. build-push v7 dropped
DOCKER_BUILD_NO_SUMMARY and DOCKER_BUILD_EXPORT_RETENTION_DAYS and
setup-buildx v4 dropped its deprecated inputs, none of which are used here.
checkout v7 refuses to check out fork PRs under pull_request_target and
workflow_run, and no workflow here triggers on either.
setup-go v6 exports GOTOOLCHAIN=local, the one change with teeth. go.mod
declares `go 1.26.5` and carries no toolchain directive, every job resolves
that same version through go-version-file, and nothing in the module graph
asks for more, so there is nothing left for the go command to fetch:
`go build ./...` passes under GOTOOLCHAIN=local. The tools installed from
outside the module are in range too -- nfpm v2.47.0 requires 1.26.4 and
govulncheck v1.6.0 requires 1.25.0 -- and golangci-lint arrives as a
prebuilt binary from upstream's install script, not through the toolchain.
The runner floor that comes with node24 (v2.327.1) is met by GitHub-hosted
runners, and every job here is ubuntu-latest.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A rebrand of this size cannot be reviewed by reading the diff. The risk is not
what was renamed but what was renamed by accident: an environment variable, a
metric, a header, a route, a storage marker or an exported symbol that some
deployment depends on. So the invariants get a machine check.
rebrand-guard extracts the identifiers a product rename must not touch - the
go.mod module path, github.com/minio/* imports, MINIO_* variables, minio_*
metric literals, x-minio-* headers, route string literals, the hardcoded grid
routes in internal/grid/manager.go, the reserved bucket root, .minio.sys
markers, arn:minio and minio:s3 policy values, every exported symbol in cmd/
and internal/, and every branded string literal - and compares them to a
checked-in baseline. Any removal or addition fails the build, in both
directions, so a future change cannot quietly drop one either.
The recorded delta against the pre-rebrand tree is: module path, 137 imports,
19 metrics, 15 storage markers, 58 policy values and all 9014 exported symbols
byte-identical; environment 437 -> 436, headers 85 -> 84, routes 331 -> 330.
Each of those three is accounted for. MINIO_CONFIG_DIR and MINIO_VERSION were a
shell local and a compose variable in test scripts, never server settings.
x-minio-common was a docker-compose extension key, never an HTTP header.
/api/health/upload was the SUBNET upload path in the deleted callhome.go.
verify-rebrand.sh covers what the identifier guard cannot see: that the retired
upstream delivery paths stay deleted, that the packaging and container files
name silo, that silo.service keeps its Conflicts and both EnvironmentFile
lines, that silo.env contains no active assignment that would shadow the legacy
file, that no active delivery surface still publishes pgsty/minio, and that no
non-test Go source contains a dl.min.io, subnet.min.io, api.min.io,
slack.min.io or play.min.io URL.
Both, plus the entrypoint argv test, run in go.yml, release.yml and
test-release.yml. test-release.yml also gets the path filters right: it was
watching minio.service, which no longer exists, so the packaging gate would
have gone silently dormant. It now watches the unit, defaults, sysusers,
lifecycle scripts, helm chart and both guards, and asserts the four-file
payload, the absence of any cross-name Provides/Obsoletes/Conflicts, and that a
container started with the legacy `minio server` argv really runs silo as PID 1.
release.yml and docker-release.yml additionally verify that the checked-out
revision matches GITHUB_SHA before building. GitHub's OIDC certificate records
GITHUB_SHA, not the ref passed to actions/checkout, so a manual dispatch from
the wrong ref would have produced provenance describing different source from
the bytes being published. They now emit SBOMs and attestations for archives,
packages and images, and docker-release verifies the release workflow's
attestation on every artifact it consumes. Images publish to pgsty/silo.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The release workflow built and pushed the container images in the same
run that created the draft, so pgsty/minio:latest moved to a build
nobody had signed or published yet. Abandoning that draft left latest
pointing at it with no way to notice, and the images were the one
artifact of a release that escaped the draft gate entirely.
Image publishing moves to its own dispatch-triggered workflow that
takes a tag and refuses to touch anything that is not a published,
non-prerelease release, and that is not the latest one - since it moves
the latest image tag, running it for an older release would silently
roll users back. It builds from the archives attached to that release,
checked against the published checksums, rather than rebuilding from
source: the image then contains the same binary the tarball does, by
construction rather than by assumption.
GoReleaser loses its dockers and docker_manifests sections along with
the QEMU, buildx and registry-login steps that only existed to serve
them, and the release job drops packages: write, which was granted for
a GHCR push that never happened.
Releasing now has one more manual step. That is the point: tagging no
longer moves docker latest, so the tag can be cut, inspected, signed
and published before anything reaches users who pull by tag.
Co-authored-by: ChatGPT <noreply@openai.com>
Co-authored-by: Claude <noreply@anthropic.com>
Feng Ruohang