Lambda cold start during ramp

On-call incidentMedium30 minAmazon style

Your task

Your API is a Lambda behind API Gateway. After a quiet overnight period, the 08:00 traffic ramp causes p99 to spike to 4–8s for the first few minutes while p50 stays at 120ms. CloudWatch shows Init Duration appearing on a large fraction of invocations during the ramp, ConcurrentExecutions climbing fast, and a few Throttles. The function loads a large ML model and opens a DB connection in its init code. No deploy happened. How do you triage and reduce the morning latency cliff?

It covers

  • Lambda
  • Serverless
  • Concurrency