Skip to content

v0.2.24

Choose a tag to compare

@robertz robertz released this 06 Nov 01:09
· 10 commits to main since this release
  • Scheduler fixes (models/Scheduler.bx):
    • A run the stall watchdog gave up on could, when it finally returned,
      mark the job idle while its replacement was still running, so the
      next tick started a third run alongside it — the overlap guard was
      off for as long as that lasted. Each run now owns the job by id and
      only the current owner can release it.
    • A job that cancelled itself during its first immediate: true run
      could beat schedule() to storing the timer's future; cancel()
      then failed, was logged as a job failure, and the job kept running.
      Cancellation now works whichever side gets there first.
    • intervalMs is validated: anything but a whole number of
      milliseconds, 1 or more, throws BoxExpress.InvalidArgument. Before,
      the JDK rejected it only after the job was registered, leaving a
      dead job in getScheduledJobs() and metrics.
    • Job ids come from an atomic counter. Two concurrent schedule()
      calls could get the same id, and the second overwrote the first,
      leaving its timer ticking with no way to cancel it.
    • Two concurrent first calls to app.schedule() could each create a
      Scheduler, and jobs registered on the one that lost were never
      reachable by getScheduledJobs(), metrics or close(). The lazy
      creation (and app.getClusterManager()'s) is now locked.
    • New options.stallThresholdMs overrides the stall threshold.
    • running (in getScheduledJobs() and the
      boxexpress_scheduler_job_running metric) counts runs actually in
      flight.
  • Cluster manager election (models/cluster/ClusterManager.bx): the
    manager claim was only written when no live manager was found and never
    renewed, so it expired every peerIdleTimeoutSeconds * 3 and went to
    whichever instance checked next, and two instances checking at once
    could both claim it and both run a clustered job. Now a live manager
    keeps the role and renews its claim on each check and heartbeat, and
    with no live manager every instance picks the same winner — the
    lowest-named live election candidate — and only it writes the claim.
    Only instances that have checked (ones with clustered jobs) are
    candidates. See "How the manager is elected" in docs/cluster.md.
  • Heartbeat interval is a third of peerIdleTimeoutSeconds with a 1s
    minimum, down from a 5s minimum that made any idle timeout under 15s
    expire live peers between heartbeats.