Replication

This page covers the current Vaylix HA model: one leader, quorum-backed writes, and follower catch-up through append and snapshot flows.

Since version: 0.5.0 introduced the current HA line; safety and test coverage continued through 0.9.0.

Caveat: The implementation is pre-1.0 and intentionally narrow. There are no linearizable follower reads, no sharding, and no distributed transactions.

What the cluster does

elects one leader
routes writes to that leader
replicates WAL-backed entries to followers
commits a write only after the configured acknowledgement condition is satisfied

Write acknowledgement

local: local durability only
majority / replica: quorum-backed acknowledgement
all: waits for every current voter

For HA use, majority is the intended setting.

Leader crash behavior

If a write was acknowledged before the leader crashed, it was already committed to a quorum and is not lost.

If a leader appended locally but crashed before quorum commit, that entry may be rolled back during leader replacement. That is expected Raft behavior.

Follower catch-up

Followers catch up through append-entry replication. If a follower is too far behind, the leader can fall back to snapshot bootstrap and then resume log application from the snapshot boundary.

Read semantics

leader reads are the default consistency boundary
follower reads are local and may be stale

Vaylix does not claim linearizable follower reads.

Operational expectations

use three voting nodes in one region
keep node IDs stable
keep peer addresses stable
route application writes to the current leader