Skip to content

v5.1.29

Choose a tag to compare

@github-actions github-actions released this 25 Sep 20:05
· 3586 commits to main since this release
68ed586

Replication: duplicate log entries from re-delivered writes

Two paths let a replicated write that a node had already applied be applied and logged again on each re-delivery. In a mesh, every re-logged copy is new log tail that peers forward and re-log in turn, so one write could grow into millions of identical audit entries under a single log key.

  • Re-delivered deletes. A delete delivered again with the same version and origin now returns without writing when the key already holds the tombstone it left. Deletes were the only write type without this identity check. (#2813, backport of #2761)
  • Writes relayed from an origin with no log on this node. Such writes are filed in the relaying peer's log, but duplicate detection only read the origin's log, so it missed every time. The write, the duplicate lookup and the retention floor now resolve the same log. (#2819)

Boot recovery: transaction-log replay no longer runs out of heap

After an unclean shutdown, replay committed only when the log key changed. A replication receiver commits every delivery of a source transaction under the origin's key, so one key can hold a very large number of separate commits. Replay staged all of them into one transaction, and the node crash-looped with an out-of-heap error on every boot. Replay now ends a transaction at each original commit's last entry, so memory is bounded by the largest single commit. (#2822, backport of #2788; fixes #2161)

Performance: user-cache refresh storm

A replicated hdb_user or hdb_role write left a flag set for the lifetime of the subscription, so every later system-database commit triggered a cluster-wide user-cache rebuild. The signal is now scoped to the commit that wrote users or roles. Each thread runs at most one user-cache scan at a time (plus one queued), and a failed replicated commit no longer raises an unhandledRejection. (#2777, backport of #2776)

Also in this release

  • Replay tests exercise the skipped-last-entry boundary against real tables; comment cleanup.

Full Changelog: v5.1.28...v5.1.29