Infrastructure & Operations › Kubernetes & Orchestration

Vertical Pod Autoscaler

Adjusting a pod's CPU and memory requests automatically.

Backend Engineer · Senior Data Engineer · Senior

Also known as: VPA, vertical pod autoscaler, resource autoscaling

The Vertical Pod Autoscaler (VPA) adjusts a pod’s CPU and memory requests and limits based on how much it actually uses. Where the Horizontal Pod Autoscaler adds or removes replicas, the VPA makes each pod the right size. It watches usage over time and recommends — or applies — better values, which helps when hand-tuned requests are too high (wasted capacity) or too low (throttling and OOM kills).

It runs in modes:

Off — only recommends; you apply the values yourself. The safest way to start.
Initial — sets requests when a pod is created, then leaves it alone.
Recreate — updates running pods by recreating them when the recommendation changes significantly.
Auto — currently recreates pods as needed, similar to Recreate, with the option to change as support evolves.

observe usage → recommend requests/limits → (in Recreate/Auto) restart pod with new values

The classic mistakes:

Combining VPA and HPA on the same metric. A VPA tuning CPU while an HPA scales on CPU utilization of that same request creates a feedback loop. Use them on different signals, or pick one.
Ignoring the restarts. Applying a new recommendation in Recreate/Auto restarts the pod. For a stateful or latency-sensitive service, that’s a disruption — pair it with enough replicas, probes and a PodDisruptionBudget.
Trusting recommendations blindly. The VPA is a helper, not an oracle. Its suggestions depend on observed load; an unusual week produces unusual values. Review important workloads.
Expecting it to fix architecture. If a service is slow because of a downstream dependency, giving it more CPU won’t help. Right-sizing is not the same as fixing the bottleneck.
Forgetting it needs a controller. The VPA targets a workload controller (like a Deployment), not bare pods.

When to use it: to stop guessing at requests and limits and keep them aligned with reality, especially across many services. Start in recommendation mode, learn what it suggests, then apply changes deliberately — and consider it part of ongoing capacity planning.