Autoscaler Arena

Controls

Load curve (drag to draw)

Sets the requests-per-second arriving over a simulated multi-hour window — the raw load driving the whole system.

Diurnal / Flash crowd / Sawtooth

One-tap presets for a smooth day/night cycle, a sudden traffic spike, or a repeating ramp — no drawing required.

Stabilization window enabled

When on, a scale-down decision only applies if load has stayed low for the FULL lookback window (real HPA behavior) — a brief dip can't shrink the fleet. When off, every recommendation applies immediately.

What to watch for

• With the stabilization window ON, watch replica count settle into a smooth curve even as load oscillates.
• Turn it OFF and replay the same load — Thrash count climbs and the replica line saws up and down.
• After a sharp load spike, watch the pod fleet: new replicas show amber (still warming up) before they count toward real capacity — latency can breach the SLO even after scale-up has already been decided.

Why it happens

The autoscaler only takes action outside a tolerance band around its target, and — critically — a scale-down decision is the MAXIMUM recommendation seen over the last several minutes, not the latest one. That means a brief dip in load can't shrink the fleet; only sustained low load for the whole window can. Strip that window away and every small dip triggers an immediate scale-down, which then has to scale back up moments later when load returns — that back-and-forth is flapping, and it's a direct, mechanical consequence of removing the lookback, not a bug.