Autoscaling as a FinOps Primary Tool
Right-sized autoscaling cuts spend more than commitment optimization. The policies are well-known; the discipline is the gap.
Why autoscaling is FinOps
Right-sized autoscaling cuts spend more than commitment optimisation does. The policies are well-known; the discipline of tuning per service is the gap.
- Hot all day. No headroom; sudden spikes degrade service; the team over-provisions to compensate.
- Cold all day. Paid-for unused capacity; the bill grows without serving traffic.
- Right-sized. Exactly the capacity needed; spike-tolerant; the savings are real.
- Beats commitments. Right-sized capacity at on-demand often beats over-provisioned capacity on commitments.
Four policy patterns
- 1. Target-tracking on CPU/memory.
- 2. Step scaling for predictable spikes.
- 3. Predictive scaling on patterned workloads.
- 4. Scheduled scaling for known time-of-day demand.
Scaling cooldown
Cooldown periods prevent thrashing. The right value is workload-specific; defaults are a starting point, not the destination.
- Default 5 minutes. Cloud defaults; reasonable for most workloads; almost never optimal.
- Too short. Oscillation; the autoscaler scales up and down repeatedly; instability.
- Too long. Lag; spike traffic hits before scaling responds; capacity arrives late.
- Tune per workload. Steady workloads tolerate longer cooldowns; bursty workloads need shorter.
Per-service tuning
One scaling policy for everything wastes 20% or more. Each service has its own demand shape; the policy must match.
- e-commerce frontend. Spiky, customer-facing; aggressive scale-up, conservative scale-down.
- Batch worker. Predictable, throughput-driven; scheduled scaling beats reactive.
- Stateful service. Slow scale-up; rebalancing cost; conservative on both directions.
- Org-wide defaults. Starting point only; per-service tuning is where the real savings live.
Antipatterns
- One scaling policy for everything. Wastes 20%+.
- Manual scaling at 3am. Unsustainable.
- No scheduled scaling for predictable demand. Pay for headroom you do not need.
What to do this week
Three moves. (1) Apply this lever to your highest-spend workload. (2) Measure the dollar impact for one month. (3) Roll the practice out to the next two services if the savings hold.