19 Releases 150 Tags

RSS Feed

• v0.8.4-beta.3 053b34c564

  Compare

  feat(cluster): join as non-voter, auto-promote/demote by reachability

  LastSkywalker released this 2026-06-20 22:57:05 +03:00 | 128 commits to main since this release

  A new node used to be added straight as a Raft voter, so a botched join (e.g.
  an advertise address that's unreachable or later goes down) could drag the
  cluster below quorum and wedge control-plane writes. Nodes now join as
  NON-voters — they replicate the full log/snapshot and serve reads, but don't
  count toward quorum — so a bad joiner can never cost the cluster its write
  availability.

  A leader-side membership manager (internal/cluster/raft/membership.go, ~10s)
  reconciles suffrage by TCP-probing each peer's advertised Raft addr:

  • promotes a non-voter to voter once it has stayed reachable >= 20s
  • demotes a voter unreachable >= 90s back to non-voter to restore quorum, but
    only while a majority is still reachable (the demote config-change must
    commit) — so the genuine 2-node 'lost a voter' case is left to the manual
    'Recover this node' action instead of churning uncommittable changes.

  The decision policy is a pure function (decideMembership) with unit tests; the
  manager reads the swappable Service so one loop survives a wipe-state
  re-activation. Adds AddNonvoter/DemoteVoter across the Service interface, the
  real Node, the disabled/swappable wrappers and the bridge test fake.

  Co-Authored-By: Claude Opus 4.8 noreply@anthropic.com
  Claude-Session: https://claude.ai/code/session_01U3LvJu9qZfspPr21TW9ptv

  Downloads

  • Source Code (ZIP)
  • Source Code (TAR.GZ)