# Cold start latency regression: pre-release checklist

Source: https://releasefailures.com/incidents/cold-start-latency-regression/
Content version: 1.0.0
Verified on: 2026-10-22
Licence: CC BY 4.0

Definition: Serverless or container cold starts become slower after a dependency or image change.
- [ ] PR-01 Measure P90 cold start duration over 100 invokes of a minimal function on the new runtime version and on the current one before switching the runtime. (test class: Load and performance tests)
- [ ] PR-02 Run the cold start benchmark at the lowest memory and CPU setting the serverless functions use in production, not with container benchmark settings. (test class: Load and performance tests)
- [ ] PR-03 Measure function init duration with and without the SDK client import on the new runtime version and fail the check when the SDK's share exceeds its share on the current version. (test class: Load and performance tests)
- [ ] PR-04 Run the init benchmark with a minimal repro that imports only the runtime's http request module, to separate runtime import cost from SDK code. (test class: Load and performance tests)
- [ ] PR-05 Benchmark preview releases of the next runtime major version in the serverless cold start scenario before the general release. (test class: Load and performance tests)
- [ ] PR-06 Report the init duration of canary cold starts as a separate metric from warm invoke duration. (test class: Canary and progressive delivery checks)
- [ ] PR-07 Halt the runtime switch when the canary's P90 init duration exceeds the baseline's by more than a set number of milliseconds. (test class: Canary and progressive delivery checks)