42% lower compute bill after requests/limits audit

Context: A B2B SaaS platform operating over 200 microservices was experiencing escalating cloud costs despite relatively stable user traffic. The platform ran entirely on a large, centralized Kubernetes cluster. Monthly compute spend hovered around $92,000, and engineers routinely added new nodes to support seemingly minor deployments.

Constraint: Development velocity was paramount. The engineering teams were shipping multiple times a day and were highly resistant to any infrastructure changes that might introduce stability risks or require developers to spend time tuning resource allocation instead of writing product code.

What we changed:

• Conducted a comprehensive audit of pod requests and limits across all deployments. We found that the vast majority of microservices were using default boilerplate requests of 500m CPU and 512Mi memory, set during initial scaffolding and never revisited.
• Deployed the Vertical Pod Autoscaler (VPA) in recommendation mode to capture baseline empirical data. The actual median usage across the cluster was just 45m CPU and 128Mi memory.
• Executed a controlled VPA dry-run, then systematically adjusted the manifests to set requests at the measured p95 usage plus a 20% safety headroom.
• Tuned the cluster scheduler configuration to prioritize aggressive bin-packing, significantly increasing workload density.

Measured result: The platform’s bin-packing efficiency improved from a highly fragmented 38% utilization to a dense 71% utilization. This allowed us to safely eliminate 34 underlying nodes from the cluster. The monthly compute bill decreased from $92,000 to $53,360, a 42% reduction, with zero impact on application performance or deployment velocity.