Skip to content

RunSignal 0.3.0

Choose a tag to compare

@github-actions github-actions released this 09 Nov 22:33
· 14 commits to main since this release
dabe1c1

Added

  • The CLI as a native executable of about 10 MB with no .NET dependency, for images and machines without .NET:
    • release assets for Linux (glibc 2.28+ and musl, x64 and arm64), macOS and Windows;
    • the container image ghcr.io/gokhanertogan/runsignal to COPY --from (/runsignal, /runsignal-musl).
  • runsignal exec: runs any command (a shell script, a crontab entry, a container in a Kubernetes CronJob) as one
    run of a workload and exports it over OTLP. Exit code 0 is a success, other codes fail with error.type
    exit_code_<n>, and stop signals are forwarded and count as cancelled. The command's exit code is returned.
  • runsignal discover kubernetes: writes a manifest from Kubernetes CronJobs, with their schedules, time zones and
    activeDeadlineSeconds, plus runsignal.io/* annotations for severity, max duration, minimum output, failure
    rate, workload name and opting out.
  • Manifest: "reporting": "per-run" for workloads whose runs are short-lived processes. missed then means no
    successful run within the schedule plus grace, and there is no not-reporting or stalled alert. On New Relic,
    per-run missed needs a run at least every 48 hours.
  • Manifest: a per-workload timeZone for its schedule.
  • Hung runs: a new runsignal.workload.active_run_age gauge in every SDK reports how long the oldest unfinished
    run has been running (0 when none is). It is reported for workloads that have maxDuration or have started a
    run. New Relic gets a hung condition and Prometheus a RunSignalHung rule, which open when an unfinished run
    passes maxDuration. Before, a run that never finished only showed up later as missed or stalled. The
    Prometheus rule reads the gauge directly, so unlike overrun it also covers limits above the largest histogram
    bucket.
  • .NET: WorkloadRun.Discard() forgets a started run without recording it. The Quartz integration uses it for
    vetoed jobs.

Fixed

  • runsignal exec no longer crashes after the command has run when it exports over HTTPS in an image without
    OpenSSL; it warns, skips the export and returns the command's exit code.
  • Alert descriptions formatted durations with the machine's culture ("1,5h" on a Turkish machine, "1.5h" in CI),
    so provision from two machines kept updating the same conditions. They are now culture-invariant.
  • Python: a second WorkloadRegistry on the same meter provider (for example one you create next to
    default_registry) silently lost its gauges, so its workloads never got missed, stalled or hung alerts.
    Registries on one provider now share their gauges.
  • Node.js: a workload in defaultRegistry that never runs reports its gauges once the SDK has started, so it gets
    its missed alert. Before, the gauges moved to the SDK's meter provider only on the first recording.
  • Java: a registered workload reports its gauges at startup, so a cron job that never fires gets its missed
    alert. WorkloadRegistry.global() does this only once GlobalOpenTelemetry is set (as the Java agent does), so
    it never fixes the global instance to a no-op.
  • Java (Spring Batch 6): when the final job repository update fails, Spring Batch skips afterJob, which left the
    run in progress. A background sweep now ends such runs as failed with error.type
    job_repository_update_failed, about 30 to 60 seconds after the job ends.
  • Java: a reactive @Scheduled method whose subscription is cancelled is recorded as cancelled, not successful.
  • Go: a run whose function calls runtime.Goexit (as t.FailNow does) fails with error.type goexit instead
    of staying in progress.
  • Rust: a registered workload reports its gauges straight away, so a scheduled workload that never runs gets its
    missed alert. Before, gauges appeared only after the first recording. Registry::global() follows the global
    meter provider until it first collects. If workloads are registered before the provider is set and record
    nothing, call the new Registry::refresh_meter_provider() once after setting it.
  • Python (Celery): a task id started twice (a redelivery, or a nested eager apply) no longer loses its first run.
    instrument() on a second registry now connects its signal handlers and records on that registry, instead of
    silently doing nothing or recording on the first one.

Changed

  • A run that is started and never ended now stays in progress, so it eventually opens hung. Before, it recorded
    nothing.