Serverless Cold Starts: Measuring What Actually Matters

Cold start numbers get quoted like a fixed constant, but the number that matters is not the vendor benchmark — it is the cold start rate under your specific traffic shape.

Why the benchmark misleads

A function that never scales past its warm pool essentially never cold-starts in production, regardless of what a synthetic single-invocation benchmark reports. Bursty, spiky traffic is what actually forces new instances, and that pattern is invisible in most published numbers.

• Provisioned concurrency trades a fixed cost for a predictable latency floor.
• Runtime choice affects cold start more than package size for most workloads.
• Traffic shape, not average request volume, determines your real cold-start rate.

What to measure instead

Track p99 latency segmented by whether the invocation was cold, using your actual production traffic — not a synthetic load test that smooths over the bursts that cause the problem in the first place.