Skip to content

v0.6.27

@lhaig lhaig tagged this 06 Sep 08:22
Raft snapshots actually happen, and the log finally truncates.

fix(cluster): apply every raft tuning value, and stop the log growing forever

  The production node had not snapshotted since 2026-08-23. Every restart
  restored a two-week-old snapshot, rebuilt derived indexes for the state as of
  that date, and warned for each message whose blob had since been collected --
  warnings that accumulate permanently, one per delete-then-GC cycle. It also
  feeds the accepted first-delivery stall (B12), which scales with log size.

  Three causes, only one of them tuning.

  snapshot_interval was parsed, defaulted, documented and set in the production
  config, and never assigned to raft.Config. A settable key with no destination
  is invisible: nothing downstream changes, so no behavioural test can see it.

  TrailingLogs was not configurable, and raft's own default of 10240 exceeds
  this project's snapshot_threshold default of 8192. Neither number is wrong
  alone, but together the threshold is inert -- a snapshot fires and then
  retains more entries than triggered it, so the log never truncates. It is now
  configurable, and an inverted pairing is refused at startup with the reason in
  the error. A node on the defaults is warned rather than refused, because
  changing those defaults would alter behaviour for existing deployments.

  The threshold was also far too high for a node delivering roughly five
  messages a day. Code defaults are unchanged; only the production config is
  tuned, to 256 with 128 trailing.

  Adds "bongo-go cluster snapshot" to force one on the leader, since a stale
  snapshot cannot otherwise be cleared without waiting for the threshold.

Known issue, not fixed here: "bongo-go cluster status" cannot authenticate
against a node with web.session_secret set. It calls http.Get unsigned against
an endpoint behind RequireInternalAuth and returns 401.
Assets 2
Loading