Kubernetes HPA Calculator & Simulator

Reproduce the exact decision the HorizontalPodAutoscaler controller makes each control-loop tick — desiredReplicas = ceil(currentReplicas × currentValue / targetValue), guarded by the --horizontal-pod-autoscaler-tolerance band, minReplicas / maxReplicas clamps, and the v2 behavior.scaleUp / scaleDown stabilization windows and per-policy rate limits (Pods or Percent per periodSeconds). Paste a load timeline — the tool ticks through every 15-second sync and shows the replica count evolve.

HPA spec

min replicas

max replicas

target metric (per pod)

metric label

tolerance (0.10 = ±10%)

sync period (s)

CPU averageUtilization=50% Requests /s target=200 Queue depth per pod=10 Memory averageUtilization=80%

Scale behavior (v2)

scaleUp

stabilizationWindow (s)

policy: Pods / period

policy: Percent / period

periodSeconds

selectPolicy

scaleDown

stabilizationWindow (s)

policy: Pods / period

policy: Percent / period

periodSeconds

selectPolicy

A policy of 0 is treated as "no limit for this policy". When both a Pods and a Percent policy are set, selectPolicy=Max (the default for scale-up) picks the larger permitted change and Min picks the smaller (the default for scale-down, so a single fat pod won't allow a big drop).

Single-tick calculator

currentReplicas

currentMetricValue (per pod)

Timeline simulator

Load timeline — one line per tick, each is total metric value (per-pod value gets divided by ready pods). Blank lines and # comments allowed.

Flat → surge → flat Sawtooth Single spike Slow ramp-down

start pods

divide load by pods?

Reference: how the HPA decides

Each sync period (default 15 s), for every ready pod, the controller reads the current metric value.
Compute the average per pod and the ratio current / target.
If |ratio − 1| ≤ tolerance (default 0.10), do nothing.
Otherwise, desired = ceil(current × ratio).
Clamp to [minReplicas, maxReplicas].
Apply the v2 behavior: scale-up uses the last scaleUp.stabilizationWindowSeconds of desireds (min), scale-down uses the max. Then apply per-policy rate caps (Pods and Percent), pick according to selectPolicy.
Emit the final desiredReplicas; the deployment does the rest.

Defaults: sync period 15 s, tolerance 0.10, scale-up unlimited + 0 s stabilization, scale-down 100%/60 s + 300 s stabilization. Kubernetes 1.29+ can tune tolerance per-HPA via spec.behavior.scaleUp.tolerance.