Cold starts still exist in 2026 - they just moved. Here is how we keep p95 and p99 honest on FaaS.
TL;DR
Trim dependencies, split cold-sensitive routes to warm pools or containers, prefetch connections, and alert on tail latency - averages lie.
When should we avoid serverless entirely?
Long-running workers, large in-memory models, and workloads with strict sub-50 ms SLOs often fit long-lived services better. Hybrid is normal - not a failure.
What is the biggest bundle mistake?
Importing entire SDKs for one helper. Tree-shake, use lightweight HTTP clients, and lazy-load optional code paths. Measure init time with real cold invokes in CI.
Refactoring a laggy API? On-call development retainers help you ship fixes without pausing roadmap work.