v0.7.1

v0.7.1

Raft-derived row versions, and one persisted owner for every shared table.

This is the first stable 0.7 release. It is not an in-place upgrade of a 0.6 or 0.7.0-beta.0 data directory. Wipe the data directory, or export and import into a fresh database. Startup does not rewrite stored rows.

Row versions

_seq is gone. Every row carries _version, sent as a decimal string:

_version = (raft_log_index << 16) | ordinal

Versions are unique and increase inside one version domain (one shared table, one owner shard, one history). They are not a single global counter, and gaps are normal.

Shared tables

Each non-partitioned SHARED table is created on one owner shard and stays there.

• Clustered default is 4 DataSharedShard groups. Standalone stays on one group and uses the same resolver.
• The owner and history are stored at CREATE TABLE. Raising shared_shards later does not move existing tables.
• Autocommit writes, live barriers, and same-group transactions follow that stored owner.
• A transaction that needs a second shared group is rejected.
• A shared live resume token must carry the exact version domain. A new subscription without from still takes a fresh snapshot.

Also since 0.7.0-beta.0

• Faster Raft apply and publisher consume, and more WebSocket clients per node
• Scalar index seeks and cooperative shutdown
• Reused native TLS clients for isolated WebSocket links
• PostgreSQL extension RPCs stay unregistered unless the extension is enabled
• DROP STORAGE removes an unreferenced filesystem directory after the catalog row is gone
• Package installs retry transient registry timeouts

What's Changed

• Replace Snowflake row sequences with a Raft-derived VersionId by @jamals86 in #390
• Keep cluster replicas consistent and let each node serve more WebSocket clients by @jamals86 in #389

Full Changelog: v0.7.0-beta.0...v0.7.1