K8s Rightsizing

K8s clusters waste 30-50%. The optimization.

Overview

K8s rightsizing matches pod requests and limits to actual usage. Node count is the easy lever; right pod sizing is what determines whether the cluster runs at 30% or 80% utilisation. K8s clusters routinely waste 30-50% of capacity.

• Clusters waste 30-50%. Per-cluster over-provisioning is the default; pods request more than they use; nodes follow.
• Per-pod requests/limits. Per-pod actual usage; the requests should match the steady-state usage, not the worst-case fear.
• VPA/HPA tuning. Per-deployment auto-scaler; VPA tunes requests, HPA tunes replicas; both have a place.
• Per-namespace quotas plus quarterly audit. Per-namespace quota prevents runaway; quarterly cluster audit catches drift.

The approach

The practical approach: per-pod request/limit tuning informed by data, VPA for requests, HPA for replicas, per-namespace quotas, quarterly audit. The team’s discipline produces matched K8s clusters.

• Per-pod requests/limits. Per-pod actual usage; tune from p95 of observed; not the architect’s estimate.
• VPA/HPA tuning. VPA recommends pod sizes; HPA scales replica count; both fed from real metrics.
• Per-namespace quotas. Per-namespace resource quota; protects against one team consuming the whole cluster.
• Per-quarter audit. Cluster review; surfaces over-provisioned and under-provisioned workloads.
• Document the policy. Per-cluster policy committed to the repo; supports operational reviews.

Why this compounds

K8s rightsizing discipline compounds across clusters and quarters. Each tuned cluster produces ongoing savings; the team’s K8s discipline grows; new workloads ship right-sized.

• Better cost efficiency. Right pod sizing matches workload; the cluster runs at design capacity, not 30%.
• Better cluster utilisation. Right cluster size matches workload; node count drops as pod sizing tightens.
• Better operational fit. Right policy matches workload; the cluster supports the workload it actually has.
• Institutional knowledge. Each pod teaches K8s patterns; the team’s capacity engineering muscle grows.

K8s rightsizing discipline is an operational discipline that pays off across years. Nova AI Ops integrates with K8s telemetry, surfaces patterns, and supports the team’s K8s discipline.