Store the container registry through object_store

[epompeii]

The container registry storage crate carried two hand-written backends: one on the AWS SDK, addressed by an S3 access point ARN, and one on the local filesystem. This replaces both with one implementation over the object_store crate, so any S3-compatible endpoint works through a bucket, an endpoint, and a region, and the local backend is object_store's filesystem store.

Changes

• One OciStorage over Arc<dyn ObjectStore> replaces OciS3Storage and OciLocalStorage. Uploads keep their session shape (chunks as objects under _uploads/, a small state.json); completion streams the chunks through a multipart upload opened directly on the final blob key and aborts on a digest mismatch, so no temporary object and no copy.
• The config section moves from plus.registry.data_store to plus.registry.storage with scheme: file | s3, mirroring the disaster recovery replica field for field. file takes an optional path; s3 takes bucket, path, endpoint, region (default auto), credentials, and chunk_size.
• Stored keys and the on-disk layout are pinned byte for byte by key layout tests, with and without a prefix, including non-ASCII and # prefixes. A multipart boundary test pins reassembly across several parts.
• The S3 client runs with no total request deadline and a 5 minute read timeout that resets on every body frame. object_store's default 30 s deadline, measured until the body finishes, cut any blob download or part upload that took longer. Two paused-time tests against a mock store pin both halves: a body that streams slowly past the old deadline completes, and a store that stalls after its headers fails at the read timeout.
• Completion sizes multipart parts so the largest allowed body fits in S3's 10,000 part limit, rather than failing at part 10,001 after streaming everything.
• The stale upload sweep is owned by the storage and stopped at shutdown, the way callbacks are drained, instead of being a detached task.
• After every upload round trip, the store is listed and only the blob may remain, which catches a chunk written outside its session prefix.
• An ignored live test (tests/s3_live.rs) runs the full flow against a scratch bucket from environment variables and cleans up after itself; it never runs in CI.
• aws-sdk-s3 and aws-credential-types leave the workspace; object_store (aws, fs) and sync_wrapper arrive. The lockfile change is additions plus the AWS tree's removal, no version bumps.
• Docs in all 9 languages describe the new section, the prefix normalization (leading, trailing, and repeated slashes dropped; control characters, non-ASCII characters, and a small reserved set such as #, %, ?, and * percent-encoded, while +, spaces, =, @, and : are kept as is), and recommend a lifecycle rule that aborts incomplete multipart uploads.

Breaking changes

• plus.registry.data_store is replaced by plus.registry.storage. A config that still carries data_store is ignored and the registry falls back to local files.
• A configured path that began or ended with a slash, or contained a percent-encoded character, stored objects under a different key prefix in the previous release; that data must be moved to the normalized prefix. Plain prefixes are unchanged.

Last of a stack of three, on top of the plus.disaster_recovery rename and the on demand backup removal.

[socket-security]

Review the following changes in direct dependencies. Learn more about Socket for GitHub.

Diff	Package	Supply Chain Security	Vulnerability	Quality	Maintenance	License
	cargo/​object_store@​0.14.2
	cargo/​base64@​0.23.1

View full report

[github-actions]

Bencher Report

Project	Bencher
Branch	u/ep/s3-r2/registry-storage
Testbed	intel-v1

Click to view all benchmark results

Benchmark	Latency	Benchmark Result microseconds (µs) (Result Δ%)	Upper Boundary microseconds (µs) (Limit %)
Adapter::Json	📈 view plot 🚷 view threshold	5.24 µs (+5.45%) Baseline: 4.97 µs	6.07 µs (86.31%)
Adapter::Magic (JSON)	📈 view plot 🚷 view threshold	5.04 µs (+4.93%) Baseline: 4.80 µs	5.76 µs (87.53%)
Adapter::Magic (Rust)	📈 view plot 🚷 view threshold	27.76 µs (+3.62%) Baseline: 26.79 µs	30.26 µs (91.75%)
Adapter::Rust	📈 view plot 🚷 view threshold	4.81 µs (+15.59%) Baseline: 4.16 µs	6.53 µs (73.74%)
Adapter::RustBench	📈 view plot 🚷 view threshold	4.82 µs (+15.89%) Baseline: 4.16 µs	6.52 µs (73.96%)

🐰 View full continuous benchmarking report in Bencher