Operator Autoscaler

2 min read

A runtime‑native autoscaler for operatorBox:es.

The Operator Autoscaler is a built‑in Orkestra subsystem that dynamically adjusts an operator’s worker count, queue depth, and resync interval based on runtime metrics, time windows, and cron expressions. It runs inside the same process as the operator and requires no external controllers, Deployments, or CRDs.

Autoscaling is fully declarative. You describe when an operator should scale and how it should scale. Orkestra evaluates these conditions at runtime and applies overrides instantly. When conditions are no longer met, Orkestra automatically restores the CRD’s declared baseline.

The autoscaler becomes the source of truth for operator performance, and Kordinator trusts it to make these adjustments throughout the operator lifecycle.

The autoscaler’s inputs

Operators are control loops, not Pods. Their performance is measured in:

  • queue pressure
  • worker utilization
  • reconcile duration
  • event throughput
  • provider error rate

These signals are already tracked per CRD inside the operatorBox. The autoscaler reads them directly from memory — no additional instrumentation required.

What the autoscaler can do

  • Scale workers up or down
  • Increase or decrease queue depth
  • Adjust resync frequency
  • Trigger scaling based on:
    • metrics (metrics.*)
    • clock windows
    • day‑of‑week
    • cron expressions
    • OR/AND logic
  • Apply overrides immediately
  • Restore baseline automatically
  • Enforce cooldown to prevent oscillation

All without restarting Orkestra.

Documentation structure

This folder contains the full autoscaler documentation:

1. Condition Engine

How autoscale conditions are evaluated, including metrics, time windows, cron expressions, and OR/AND semantics.

2. Metrics

The autoscaler’s metric namespace (metrics.*), how metrics are collected, and what each metric means.

3. Runtime Behavior

How the autoscaler loop runs inside Kordinator, how overrides are applied, and how baseline restoration works.

4. YAML Reference

The complete schema for operatorBox.autoscale, including examples and field‑level explanations.

5. Scenarios

Real‑world autoscaling patterns: business hours, nightly batch windows, traffic‑based scaling, weekend scale‑down, and more.

Mental model

Think of the autoscaler as a mini‑controller inside Orkestra:

  • The CRD defines the steady‑state baseline.
  • The autoscaler decides when to override it.
  • Kordinator applies the decision live.

This separation mirrors Kubernetes (Deployment → HPA → Kubelet), but applied to operators themselves.

Try it

ork init --pack advanced
cd 12-autoscale/01-without-autoscaler

Follow the README — it starts with baseline behavior (no autoscaler) and adds progressively more complex autoscaling across five examples.