Skip to content

autoscaling.md

Latest commit

 

History

History
38 lines (31 loc) · 1.32 KB

autoscaling.md

File metadata and controls

38 lines (31 loc) · 1.32 KB

Autoscaling

HPA scaling signal: temporal_task_queue_backlog — a Prometheus recording rule that sums approximate_backlog_count across all partitions and versions. This global metric is responsive even with short-lived workflows, making it well-suited for demo scenarios.

A per-version metric (temporal_worker_slots_used_by_version) is also available as an external metric for production use with longer-running workloads.

Pipeline (backlog): Temporal Server (Prometheus scrape) → PodMonitor → Prometheus → recording rule → prometheus-adapter (external metric) → HPA.

Pipeline (slots): Worker SDK (Micrometer) → OTLP → OTel Collector → Prometheus remote write → recording rule → prometheus-adapter (external metric).

Version-identifying tags are added as Micrometer common tags in the K8s profile of application.yaml:

  • temporal.worker.deployment.name
  • temporal.worker.build.id
  • temporal.namespace

These translate to underscore-separated Prometheus labels via UnderscoreEscapingWithSuffixes, matching the labels the Worker Controller auto-injects into each versioned HPA's matchLabels.

Key files:

  • k8s/worker/worker-resource-template.yaml — HPA template with temporal_task_queue_backlog
  • worker/src/main/resources/application.yaml — version tags (K8s profile, second YAML document)