[dimitri-yatsenko]

Closes #1560. Supersedes #1608.

Summary

The test suite and the docker-compose dev environment now use SeaweedFS as the S3-compatible object store, in place of MinIO. MinIO has stopped publishing public images and archived its source: minio/minio is gone from Docker Hub, quay.io/minio/minio requires authentication, and the Chainguard build adopted in #1559 depends on a third party rebuilding software that no longer gets upstream fixes.

Only the test and development setup changes; no library code is touched. DataJoint's support for MinIO as a user's storage backend (protocol: "minio" in migrate.py, and any S3-compatible endpoint generally) is unchanged.

Changes

File	Change
tests/conftest.py	minio_container → s3_container, running chrislusf/seaweedfs:4.48 through a generic testcontainers DockerContainer, with a readiness wait on the S3 gateway. Image override DJ_TEST_MINIO_IMAGE → DJ_TEST_S3_IMAGE. Fixture minio_client → s3_client. Marker requires_minio → requires_s3. External-container default S3_ENDPOINT → localhost:8333.
tests/integration/test_{npy_codec,attach,codecs}.py	Use the renamed s3_client fixture.
docker-compose.yaml	The minio service becomes s3, on the same image as the fixture, port 8333, with a healthcheck; app depends on it. The previous minio/minio image no longer pulls, so docker compose up was already broken.
pyproject.toml, pixi.lock	Drop the unused minio extra from testcontainers (which pulled in the minio client and pycryptodome); rename the marker; services-up starts db s3.
CONTRIBUTING.md	Updated service names, commands, and endpoint.

Credentials (datajoint/datajoint) and the bucket name (datajoint.test) are unchanged.

Why SeaweedFS

#1560 set the bar: DataJoint reaches object storage through s3fs, which emulates directories over S3 listing, and garbage collection decides what is orphaned by walking that emulation. A server that lists differently would not error; it would make GC misclassify files while test_gc still passed. That rules out lenient mocks (moto, s3mock) as the default. SeaweedFS is a real object store (Apache-2.0, public image, monthly releases), and it was tested first; Garage was the fallback and was not needed.

Testing

Listing fidelity. A probe built the same tree on Chainguard MinIO and on SeaweedFS through s3fs: nested prefixes, a zero-byte directory marker, a marker that also has children, an empty object, and a missing path. It recorded isdir, exists, ls, find (with and without directories), walk, du, isfile on the marker, and the tree after a recursive rm. The outputs were identical.

Suites.

Run	Result
Object-storage suites (test_gc, test_object, test_npy_codec, test_codecs, test_attach) on MinIO, baseline	124 passed, 41s
Same suites on SeaweedFS, cold pull with the image removed first	124 passed, 30s
Same suites against docker compose up s3 with DJ_USE_EXTERNAL_CONTAINERS=1	124 passed; the service reports healthy
Full suite on SeaweedFS	1203 passed, 14 skipped
-m requires_s3	selects 39 tests, as requires_minio did

Auth. Credentials are enforced: a request with a wrong secret is rejected.

Notes for reviewers

• Anyone with a local DJ_TEST_MINIO_IMAGE, or with -m "not requires_minio" in their own scripts, needs the new names.
• For anyone running external containers by hand, the endpoint moves from port 9000 to 8333.