Kubernetes Cost Optimization: Saving $200K/Year for a SaaS Client

The Problem

Our client, a B2B SaaS company with 500+ enterprise customers, was spending $45K/month on their Kubernetes infrastructure. Growth was strong, but cloud costs were growing faster than revenue. At the current trajectory, infrastructure would consume 40% of their revenue within 18 months.

They asked us to cut costs by 30% without impacting performance or reliability. We delivered 37%.

Cost Assessment

Before optimizing, we spent two weeks measuring everything:

• Resource utilization: Average CPU utilization was 12%. Memory utilization was 23%. Massive waste.
• Cost allocation: 60% of spend was compute, 25% storage, 15% networking
• Over-provisioning: 73% of pods had resource requests 3x higher than actual usage
• Idle resources: Development and staging environments ran 24/7 despite being used only during business hours
• Storage waste: 2TB of unused PVCs from deleted deployments

The assessment alone identified $15K/month in easy wins.

Right-Sizing Workloads

We implemented a data-driven right-sizing strategy:

• VPA recommendations: Deployed Vertical Pod Autoscaler in recommendation mode for two weeks, then applied suggestions
• Resource quotas: Set namespace-level quotas to prevent over-requesting
• QoS classes: Classified workloads as Guaranteed (databases), Burstable (APIs), and BestEffort (batch jobs)
• Node pools: Created specialized node pools — compute-optimized for APIs, memory-optimized for caches, spot for batch

Result: 40% reduction in requested resources with zero performance impact.

Autoscaling Strategies

We implemented multi-layer autoscaling:

• HPA with custom metrics: Scaled based on request latency and queue depth, not just CPU
• KEDA for event-driven scaling: Scaled workers based on message queue length — zero pods when idle
• Cluster autoscaler tuning: Reduced scale-down delay from 10 minutes to 2 minutes, enabled scale-to-zero for dev/staging
• Scheduled scaling: Pre-scaled for known traffic patterns (Monday morning surge, end-of-month reporting)
• CronJob for non-prod: Shut down dev/staging environments outside business hours (6PM-8AM, weekends)

Autoscaling alone saved $8K/month.

Spot Instance Strategy

Spot instances were the biggest cost lever:

• Stateless workloads on spot: All API servers and workers run on spot instances (70% cost savings)
• Multi-AZ, multi-instance-type: Spread across 8 instance types and 3 AZs for availability
• Graceful shutdown: All services handle SIGTERM with a 30-second drain period
• Spot fallback: Automatic fallback to on-demand if spot capacity is unavailable
• PDB (Pod Disruption Budgets): Ensure minimum replica count during spot interruptions

We achieved 95% spot coverage for stateless workloads with zero customer-facing interruptions.

Final Results

After 8 weeks of optimization:

• Monthly cloud spend: $45K → $28.3K (37% reduction, $200K annual savings)
• CPU utilization: 12% → 45% average
• Memory utilization: 23% → 58% average
• P99 latency: Unchanged (actually improved slightly due to right-sized containers)
• Availability: 99.98% maintained
• Developer experience: Improved — dev environments spin up 3x faster with right-sized resources

The optimization paid for itself in the first month. The client now runs quarterly cost reviews using the dashboards and processes we put in place.