What causes cold starts in serverless functions and how do you deal with them?

A cold start happens when a request arrives with no warm execution environment available, so the platform has to provision one, download and unpack your code, start the runtime, and run your initialization before handling the request. Reduce it by shrinking the package, choosing a lighter runtime, moving heavy setup out of the request path, and reusing connections. For latency sensitive paths, use provisioned concurrency rather than warming hacks.

A cold start is the platform building a new execution environment because there is not a warm one free. That means provisioning it, pulling and unpacking your artifact, booting the runtime, and running your initialization code before your handler is even called. The parts I control are the artifact size and the initialization, so I trim dependencies hard, because a fat package adds real time to every cold start, and I move things like configuration loads and client construction outside the handler so they run once per environment rather than once per request and get reused on warm invocations. Runtime choice matters too; a lightweight runtime starts far quicker than a heavyweight virtual machine unless I use the platform's snapshot feature. When the path is user facing and the tail latency matters, I stop optimizing and buy provisioned concurrency for the expected baseline, because that keeps environments initialized. What I do not do any more is ping functions on a timer to keep them warm; it is unreliable at any real concurrency. For queue driven or batch work I usually just accept it.