Consistency
Strong consistency via Raft for cluster metadata.
Replicated (quorum write)
- Membership and leadership
- Locks and leases (including presence)
The leader is the only writer. Followers and observers forward mutations they cannot commit (except observer locks, which are rejected).
Not replicated
- Application data, files, large datasets
- Custom event payloads
Custom events are 1-hop gossip. A partition loses those signals. Membership and locks on the log are not lost that way.
Putting user payloads on the Raft log would make every publish a quorum write and bloat snapshots. That is a non-goal.