Most Kubernetes cost-optimization advice focuses on rightsizing pods. That helps, but the bigger wins are usually structural — cluster autoscaler configuration, node pool strategy, and how you handle bursty batch workloads.
Beyond pod rightsizing
Rightsizing requests and limits based on actual usage (not guesses) is table stakes, but it typically saves 10-20% — the bigger wins come from node pool strategy and workload scheduling.
Running a mix of on-demand and spot/preemptible node pools, with pod disruption budgets and taints/tolerations routing fault-tolerant batch workloads to the spot pool, routinely saves 40-60% on compute for those workloads.
Cluster autoscaler tuning
Default autoscaler settings are conservative by design. Tuning scale-down delay and utilization thresholds to match your actual traffic patterns — rather than leaving headroom “just in case” — is free money left on the table in most clusters we review.
Bin-packing and pod density
Two clusters running identical workloads can have wildly different costs based purely on bin-packing efficiency. Using a scheduler that considers both CPU and memory fragmentation (not just raw utilization) can meaningfully increase pod density per node.
Leave a Reply