Skip to content

CHANGELOG.md

Latest commit

 

History

History
76 lines (65 loc) · 3.68 KB

CHANGELOG.md

File metadata and controls

76 lines (65 loc) · 3.68 KB
Raw

Changelog

All notable changes to zs3 are documented here. The format follows Keep a Changelog, and versioning follows Semantic Versioning.

[0.1.0] - 2026-08-09

First release. Distributed mode now replicates across nodes: writes leave the node that receives them, reads fall back to peers, and a gossip round repairs the mesh after restarts and partitions.

Added

  • Distributed namespace replication. PUT/DELETE/multipart-complete push the raw metadata entry (inline data and tombstones included) to all peers via POST /_zs3/meta. Bucket create/delete propagate via POST /_zs3/bucket[_delete]. Cross-node GET/LIST after an acknowledged write is immediately consistent.
  • Read fallback. A GET/HEAD miss falls back to POST /_zs3/meta_get. Blob reads sweep all known peers when provider records are missing.
  • Bootstrap sync. Joining nodes pull the full index (GET /_zs3/index) and discover the bootstrap peer's peers, handshaking so links are bidirectional.
  • Background replication worker. Blob replication, DHT announces, and bucket-op propagation run on a PushWorker thread off the write path, so copying to peers no longer adds to write latency. The ~1KB metadata push stays synchronous.
  • Peer gossip. A --gossip-interval-ms round (default 30s) pings random peers, refreshing last_seen and pulling peer lists, so a node that restarts with an empty routing table is rebuilt without re-bootstrapping.
  • Address gossip. /_zs3/peers entries now carry a full "addr":"ip:port" field. The port-only form is kept as a fallback.
  • Clean 4xx errors for filesystem-size limits. Keys that can't fit a filesystem (NAME_MAX per component, PATH_MAX total) return 400 KeyTooLong in both standalone and distributed modes instead of a misleading 500 from ENAMETOOLONG.
  • 431 for oversized headers, 400 for oversized bodies. Request headers over 8KB get 431. Bodies over 5GB get 400. Neither is truncated or dropped silently.

Fixed

  • Bucket resurrection race. Bucket create/delete are broadcast by different nodes at different times, so a late-arriving create could resurrect an already-deleted bucket on replicas (flaky HEAD <bucket> in the replication suite). A per-bucket last-write-wins registry (BucketOps, persisted in data_dir/.buckets/) keeps create/delete timestamps. Peers apply only newer ops, and on a tie the delete wins, matching the object-metadata tombstone convention.
  • fetchFromPeer EINVAL panic on macOS. A single ~5GB read() is rejected by macOS with EINVAL and panics std.Io. Peer responses now read in 16KB chunks.
  • Hung peers can no longer wedge the event loop. Peer sockets get 5s send/receive timeouts, so a hung peer can't deadlock two nodes mid-request.
  • Unauthenticated /_zs3/ endpoints now validate bucket/key names.
  • HeadObject for directory prefixes. Directories answer HEAD instead of a miss.

Docs

  • Position zs3 as "SQLite for objects": new docs/vision.md, docs/replace-minio.md, docs/deployment.md update, and scripts/verify-clients.sh / scripts/verify_boto3.py for reproducing the aws-cli / boto3 / rclone compatibility check.

Tests

  • New test_replication.py: a four-node suite (113 checks) covering cross-node reads, LWW overwrites, delete-then-recreate, multipart, pagination, restart catch-up, origin-node death, gossip mesh repair after a routing-table wipe, and peer protocol validation. Unit tests cover the meta-entry parser.
  • Distributed boto3 suite: 71 → 72 passing.
  • Regression tests for the 400 KeyTooLong / 431 / 400 error paths in test_client.py.