What’s the fastest way to prove ROI on performance work?

Instrument RUM and correlate LCP/TTFB with conversion by device/geo. Then edge-cache the hottest unauthenticated endpoints and kill the top 5 slow queries. You’ll usually see measurable conversion lift within a week and reduced origin cost from offload.

How do we pick a cost SLO?

Start with historical spend divided by requests over a stable period—say $0.70 per 1k requests. Set an initial SLO 10–20% lower to force optimization. Track egress separately; it often hides big wins via caching and image compression.

Is switching to Graviton worth the churn?

Often yes for stateless services. We regularly see 20–40% better price/perf. Validate with a canary pool under production traffic. For JVM stacks, ensure you’re on recent JDKs and tune GC; for Go/Node, changes are typically minimal.

Can we autoscale on RPS in Kubernetes without vendor lock-in?

Yes. Use Prometheus metrics and the Prometheus Adapter to expose custom metrics to HPA. Scale on requests_per_second or queue depth instead of CPU. Keep CPU as a fallback to avoid sudden drops.

How do we keep optimization from becoming a one-off project?

Make efficiency SLIs first-class: put $/req and RPS/vCPU on the same dashboard as p95. Review in a weekly perf standup with product and finance. Tie promotions and feature rollouts to error budgets and performance guardrails.

Performance-optimization · Nov 3, 2025 · 10 minute read

Stop Buying CPUs for Bad Code: A Pragmatic Framework to Balance Performance and Cloud Spend

If your p95 is creeping up and your AWS bill is doing the same, you don’t need a new instance type—you need a resource optimization framework that ties user experience to dollars.

Back to all posts

Stop Buying CPUs for Bad Code: A Pragmatic Framework to Balance Performance and Cloud Spend

Key takeaways

Implementation checklist