19 Releases 150 Tags

RSS Feed

• v0.11.0 29a19f7c64

  Compare

  v0.11.0 — a node cut off from its cluster says so, and offers the way back

  LastSkywalker released this 2026-09-12 21:25:36 +03:00 | 58 commits to main since this release

  A node whose advertised Raft address stops reaching it used to be invisible:
  peers dialled a dead address, so it heard no leader and could commit nothing,
  while its outbound metric stream kept its card green and its replicated record
  silently froze. This release recognises that state, marks the machine on every
  dashboard, and offers a one-click move back into the cluster.

  • Deaf-node detection: after two minutes without a leader the node asks its
    peers and compares its advertised address with the ones the machine holds.
    Silence from every peer yields no verdict — an outage must not read as a
    misconfiguration.
  • The fault travels on the metric stream (the one channel such a node still
    has), so the hub, the peers and the node itself all mark the same machine.
    The UI stays a marker plus one line; the click leads to the settings of the
    machine that needs the fix, wherever you look from.
  • A Fix button moves the node: it persists the corrected address, restarts the
    cluster layer on it keeping all data, and asks a peer to update the
    membership — the part only a leader can do and a cut-off node cannot reach.
    The request is HMAC-signed with the cluster secret.
  • While the fault lasts the node stops burning work: the metric stream drops to
    a liveness cadence and the failing record write backs off, instead of a
    consensus attempt and a log line every tick.

  Also fixed: a node that failed to publish its URL at startup (the normal case —
  it runs before an admin adds it) recorded the attempt as done and stayed absent
  from the peer catalog for ten minutes, so no peer streamed metrics to it and
  every other machine's card sat empty on its dashboard.

  Downloads

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