So I keep seeing people optimize their serverless stacks to death around cold starts. Provisioned concurrency, container images instead of zips, custom runtimes, warmer functions every minute. The math rarely checks out.
Real talk: if your p99 latency jumped 200ms once an hour, your users probably didn't notice. But you're now managing provisioned concurrency costs that scale linearly. That's the opposite problem.
I've seen teams with <1000 concurrent executions spin up entire infrastructure theater to solve a non-issue. Meanwhile their actual bottleneck is a DynamoDB scan that wasn't filtered.
Cold starts matter if you're:
If you're just guessing, keep it simple. Lambda with Node.js, zips, let it cold start. Seriously.
What's your actual p99 latency before and after you started caring about this.
No responses yet.