Skip to content

v0.6.24

@lhaig lhaig tagged this 30 Aug 14:08
Two data races in the Raft FSM, and the per-user data leak behind the
ghost-mailbox incident.

fix(cluster): synchronise FSM dependency injection against the apply loop

  Agent.Run injects the search indexer, blob syncer, notifier and derived-index
  queue into the FSM after handing it to Raft, so the apply goroutine read those
  fields while Run was still writing them. This fires on any restart that
  applies Raft entries during startup - i.e. every restart with a non-empty log.
  StartIndexer's check-then-assign was also a double-start window: two callers
  could each create a queue, orphaning the first worker. All four fields are now
  guarded and read through accessors.

fix(admin): reclaim per-user data left behind by account deletion

  DeleteUser removed /user/{id}/* and the users row but neither the per-user
  metadata database nor the search index, so every deleted account leaked both.
  That leak is what let delivery keep succeeding into a deleted account's
  mailbox. Adds the cascade, plus an orphan-GC sweep that reclaims /user/{id}/
  trees with no users row and no profile - which also unpins the blobs those
  keys were holding.

chore(ci): stop running every workflow twice per pull request
Assets 2
Loading