@absolutejs/autoscaler

Horizontal-scaling policy used by the hosted AbsoluteJS.ai platform and available to any Bun control plane.

@absolutejs/autoscaler owns the decision half of an autoscaler. The actuator half — actually provisioning a VM, draining a pod, killing a process — lives in your control plane via an injected Actuator, so the substrate stays cloud-agnostic and runtime-agnostic.

The same substrate fits:

• a 10-VM fleet across DigitalOcean / Hetzner / Linode
• a 10000-isolate fleet on one box via @absolutejs/isolated-jsc
• a pod-per-tenant cluster on Kubernetes

The actuator defines what "instance" means.

Loop

read capacity + signals  →  restore hard min/max bounds first  →
otherwise compare the combined score to thresholds  →
if past threshold & cooldown elapsed  →  ask actuator to spawn/drain

API

import {
  createAutoscaler,
  createPolicy,
  ratioSignal,
} from "@absolutejs/autoscaler";

const scaler = createAutoscaler({
  policy: createPolicy({
    min: 1,
    max: 20,
    scaleUp: { threshold: 0.75, cooldownMs: 60_000, step: 1 },
    scaleDown: { threshold: 0.3, cooldownMs: 300_000, step: 1 },
  }),
  signals: [
    ratioSignal("cpu", 0.8, async () => await meter.cpuUtilization()),
    ratioSignal("queue", 100, async () => await queue.depth(), {
      observedKey: "depth",
    }),
    ratioSignal("latencyP95", 200, async () => await metrics.p95()),
  ],
  combine: "max", // worst pressure wins. or 'avg', or a custom fn
  actuator: {
    list: () => fleet.list(),
    spawn: () => fleet.provision(),
    drain: (id) => loadBalancer.remove(id),
    terminate: (id) => fleet.destroy(id),
  },
  audit: broker, // optional; emits autoscaler.scale.up etc.
  intervalMs: 30_000,
});

scaler.start();
// fires every 30s

const reviewedPlan = await scaler.evaluate();
await scaler.applyDecision(reviewedPlan, { maxAgeMs: 300_000 });
// applies that exact plan only while its capacity precondition still holds

const oneShot = await scaler.step();
// { action: 'scale-up' | 'scale-down' | 'hold', score, currentCount,
//   desiredCount, reason, readings: [...], at }

Signals

A Signal is { name, read: () => SignalReading | Promise<SignalReading> }.

A SignalReading is { score, observed? }. score is normalized: 1.0 = "at the scale-up target." Going past 1.0 represents over- target pressure.

The bundled ratioSignal(name, target, read) helper builds the canonical "observed / target" shape for CPU utilization, memory pressure, queue depth, latency, etc.

A signal that throws is captured as { failed: true, error } in the decision readings and excluded from the combined score. The loop never breaks because one metric source is offline.

Combine strategies

• 'max' (default) — worst pressure wins. The safe choice for elasticity: any one over-stressed dimension scales the fleet.
• 'avg' — weighted mean using each signal's weight. Zero- weight signals are ignored.
• Custom function — (readings) => number for arbitrary composition (latency-weighted CPU, percentile-of-percentile, …).

Cooldowns

scaleUp and scaleDown have independent cooldown timers — a fleet that just scaled up can still scale down moments later if the load drops. Default cooldownMs is 60 seconds.

Audit trail

When an audit broker is supplied, every applied decision emits:

• autoscaler.scale.up — with { score, currentCount, desiredCount, reason }
• autoscaler.scale.down
• autoscaler.hold

All three are emitted (even hold) so the audit trail tells the full story of why the fleet is where it is. Broker failures are isolated; they bump the metrics().errors counter but never break the loop.

License

BSL-1.1 with named carveout against hosted autoscaling / elasticity SaaS (Karpenter SaaS offerings, Spot.io, CAST AI, PerfectScale, Densify, Granulate). See LICENSE. Change date: 2030-05-31 → Apache 2.0.