Explain the concept of cold starts in AWS Lambda and how they can affect performance.

Example Answer

A cold start includes preparing a new execution environment and running initialization before the handler can process a request. It can happen during scale-out or after an environment is replaced, not only after a function has been idle. I would measure initialization separately from handler and downstream latency, then reduce expensive imports and initialization work. For a latency-sensitive API, provisioned concurrency can keep allocated environments initialized, but requests beyond that capacity can still use on-demand environments. Scheduled warm-up calls do not guarantee that later requests use the same environment.

Investigation example

Compare cold and warm latency distributions, initialization duration, package contents, and network calls performed during initialization. A large model loaded at module scope may reduce repeated warm-handler work while making initialization slower; that tradeoff should be measured.

VPC networking uses managed Hyperplane ENIs that can be shared and reused. Do not teach the old model of creating a fresh ENI for every cold invocation. Test a burst above prepared capacity and an initialization failure, not just one warm request.

Practice follow-up

When would the additional cost of provisioned concurrency be justified by the user-facing latency target?

Sources

Lambda execution environment lifecycle; Lambda provisioned concurrency; Lambda Hyperplane networking.