ServiceRadar v1.4.66

Add-on rollouts no longer wedge: a superseded rollout finishes
its source, a candidate is no longer failed on the health report
of the version it replaces, and the agent restarts an add-on
whose RPC server died while its process stayed up. Existing
anomaly settings pick up the per-class drift mode defaults.

Whats New

1.4.66

• Add-on rollouts: a rollout superseded because the fleet already
  runs the candidate now finishes like a completed rollout. It
  promotes the source when the source is behind the candidate,
  clears per-assignment overrides and releases the target slots.
  Before, the profile stayed on the old package, every retry
  failed on addon_rollout_targets_one_active_target_index, and
  the source could never roll out again. Reconcile repairs
  rollouts already stuck this way (PR #449, issue #445).
• Add-on rollouts: a candidate is failed only on a health report
  for the candidate version. An unhealthy report from the version
  being replaced no longer rolls the upgrade back; an agent that
  never applies the candidate still fails as
  candidate_health_timeout (PR #449, issue #430).
• Agent: the add-on supervisor restarts an add-on whose process is
  alive but whose RPC server is gone, instead of reporting it
  unhealthy until someone restarts the agent (PR #449).
• Anomaly: migration 20260914120000 fills the per-metric-class
  drift_mode defaults into existing anomaly settings where a
  class has none (cpu, memory and interface
  deseasonalized_only; disk, icmp and other off). Values an
  operator already set are left untouched (PR #452).
• Jobs: 49 Oban queries looked up a worker's jobs by
  to_string(Module), which keeps the Elixir. prefix Oban strips
  from the stored name, so none of them matched. "Already
  scheduled" guards now work: the sweep coverage rollup stops
  adding a daily chain every hour and the Bumblebee catalog
  refresh stops adding one per coordinator start, and both cancel
  existing duplicates. Composite check cancel, the GeoLite and
  IPinfo "download now" promotion and the IP enrichment stuck-job
  reaper now take effect. A Credo check rejects the old form
  (PR #455, issue #448).

Upgrade Notes

• New migration 20260914120000. The chart default for
  core.migrations.expectedVersion moves to 20260914120000; a
  deployment that pins that value in its own values file (manual
  Helm installs) must bump it in the same upgrade.
• The add-on supervisor fix ships in the agent binary, so hosts
  pick it up through the normal agent release rollout.
• On first start, the sweep coverage rollup and Bumblebee
  catalog refresh cancel duplicate scheduled jobs down to one
  chain each and log "Cancelled duplicate pending Oban jobs" once.
  IP enrichment refresh runs now stop after half of
  stale_executing_minutes (30 minutes by default).