19 Releases 150 Tags

RSS Feed

• v0.11.3-beta.0 9b5e56cce0

  Compare

  fix(hosts): a node showed itself twice — fold the stale copy of its own machine back into id=1

  LastSkywalker released this 2026-09-13 19:33:00 +03:00 | 56 commits to main since this release

  After the previous fix the cluster agreed on one row per machine, but the moved
  machine's OWN dashboard still carried a second, frozen card for itself. The two
  rows were the same machine by SMBIOS UUID, split only by MAC.

  id=1 is written locally and always carries the MAC the agent has right now. When
  the cluster moved that MAC onto its copy of the same machine, the update was
  deflected — mac_address is unique and id=1 already held it — so the copy kept the
  old MAC AND the external_id. From then on it won every external_id lookup in
  UpsertConnectorHost, id=1 never received the topology, and each new connector
  poll refreshed the wrong row. Nothing in the flow could ever undo that: the
  lookup that picks the wrong row is the same one that would have to fix it.

  The upsert now recognises a resolved row as a stale copy of this node's own
  machine and folds it into id=1, dropping it. The gate is deliberately the
  strongest evidence available — the two rows sharing a non-empty stable identity
  (SMBIOS UUID or machine-id). Not a MAC: Docker-bridge agents on different
  machines derive identical 02:42:… ones, and a mistaken fold here would swallow a
  real machine's row. A genuinely remote origin is excluded for the same reason the
  cross-site guard above it exists.

  Also drops a misleading branch from the previous commit. The FSM treated a gap
  above its watermark as proof that a snapshot had been installed and
  fast-forwarded past it, logging so on a live cluster every few minutes: raft
  keeps configuration and no-op entries out of the FSM entirely, so the indexes it
  hands us legitimately skip. The watermark never needed to be contiguous with the
  log — a restore realigns it explicitly — and the claim it printed was simply
  untrue.

  Verified against the three real nodes: the leader and the LXC had already
  converged to one row per machine after v0.11.1, with live watermarks; the NAS
  still held id=1 (MAC …:00:02, no topology) and a copy on …:00:03 holding
  external_id …/qemu/101, both carrying UUID 4b23b4fd — exactly the split this
  folds. Covered by tests for both the fold and the two-machines-must-stay-apart
  case.

  Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com

  Downloads

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