Total cold start time is not enough — you need to know which import, module, or connection is responsible for each millisecond. LLRT benchmarks, X-Ray init tracing, and lazy loading reveal the exact bottleneck.

Lambda container images up to 10GB unlock PyTorch, large binaries, and custom runtimes impossible with ZIP. But ECR layer caching changes cold start behavior in ways most teams misunderstand. Here is when containers win.

Step Functions Express Workflows run at 100K+ executions/second at $0.000001 per state transition — 25,000x cheaper than Standard at scale. The right orchestration layer for high-volume Lambda event processing and ETL pipelines.

Provisioned Concurrency eliminates cold starts at $14/month per unit. SnapStart gives 96% of that benefit for Java at zero cost. Reserved Concurrency prevents throttling but does not warm instances. Here is exactly when to use each.

Lambda VPC cold starts used to add 10 seconds. AWS fixed this with Hyperplane ENIs in 2019 — but most teams still configure VPC in ways that break the optimization. Here is what actually works.

Lambda Extensions are sidecar processes that run in the same execution environment as your function. They power Datadog, New Relic, and HashiCorp Vault integrations. Here is how to build your own.