Release Incident Playbook

Cold start latency regression: symptoms, cause, and the test that catches it

Serverless or container cold starts become slower after a dependency or image change.

Cold start latency regression is documented by 2 accepted public reports from 2 organisations, 2 of them published in the last 5 years.

  • Likelihood: low (2 public reports in 5 years)
  • Impact: low (2 of 2 reports: degraded)

2 public reports · verified 2026-10-22

Symptoms

User-visible

  • The first invoke of a hello world serverless function, where the managed language runtime starts, was 5 to 10 percent slower on the new runtime major version than on the version two majors earlier.[1]
  • Serverless cold starts were at least 50 ms slower after functions that use the cloud platform's JavaScript SDK moved to a newer JavaScript runtime major version.[2]

Monitoring-visible

  • P90 total duration over 100 cold start invokes was about 730 ms on the new managed runtime version, against about 680 ms and about 671 ms on the two previous versions.[1]
  • Runtime event captures showed more types loaded at startup on the new managed runtime version.[1]
  • Function init duration with the JavaScript SDK was about 255 ms on the newer JavaScript runtime against about 204 ms on the previous one, with no difference when the SDK was absent.[2]

Root causes

  1. stage: third-party Many types were added to the new managed runtime version, and the low-CPU serverless startup scenario is sensitive to the extra type loading cost.[1]
  2. stage: third-party The JavaScript SDK imported the runtime's http request and agent modules at startup, and on the newer JavaScript runtime loading those modules adds init time even when they are not used.[2]

Public examples

  1. [1]Microsoft (.NET runtime), 2025-10-01. An AWS Lambda engineer measured a 5 to 10 percent slower first invoke on .NET 10 than on .NET 8 in Lambda; .NET team members attributed it to extra type loading cost from types added in .NET 10 and backported a lazy initialisation fix to .NET 10 servicing. primary Detecting .NET 10 Regression in startup performance affecting AWS Lambda
  2. [2]Amazon Web Services (AWS SDK for JavaScript v3), 2025-03-02. After moving Lambda functions that use the AWS SDK for JavaScript v3 from Node 20 to Node 22, init duration rose by about 50 ms because the SDK eagerly imported node:http; the SDK team shipped lazy loading in @smithy/node-http-handler 4.5.0. primary Lambda Node 22 coldstart latency regression - defer loading of node:http

Test classes that catch it

Test classStageLimitationBasis
Load and performance testspre-deployA startup benchmark run with container settings showed the new managed runtime version as faster, so a performance test misses the regression unless it reproduces the low-CPU serverless cold start.by design analysis
Canary and progressive delivery checkspost-deployA canary that compares request duration without separating init duration blends the cold start increase from the SDK's startup import into warm invokes that do not pay it.by design analysis

Risk

  • Likelihood: low (2 public reports in 5 years)
  • Impact: low (2 of 2 reports: degraded)
Counts behind the grades
Accepted public reports2
Published in the last 5 years2
Impact stated: degraded2

How the grades are computed

Pre-release checklist

  • PR-01
  • PR-02
  • PR-03
  • PR-04
  • PR-05
  • PR-06
  • PR-07

First 30 minutes

  • FM-01
  • FM-02
  • FM-03
  • FM-04
  • FM-05
  • FM-06
  • FM-07

Download the checklist

What catches this: catalogue entries

Listed because their own documentation or site names load and performance tests, canary and progressive delivery checks on the verification date. Alphabetical. Not a recommendation.

Tools whose documentation names these test classes

Providers whose site names these test classes

Only on this page

Cold start latency regression starts in the third-party stage in report [1] and report [2], not in application code. Neither team changed its own function code. In report [1], the slowdown came from types added to a new managed runtime version [1]. In report [2], it came from the SDK's startup import of the http module on a newer JavaScript runtime [2]. A code review of the function shows no diff that explains the slower cold start.

In report [1], the runtime maintainers' own startup measurements used container settings in which the new version was faster [1]. The regression was found by the serverless platform's performance testing before the new runtime became a managed option [1]. Both fixes shipped upstream as lazy loading: a runtime servicing release in report [1] and an http handler release in report [2].

Questions

What causes cold start latency regression?

Both reports name a third-party cause. In report [1], types added to a new managed runtime version raised type loading cost at startup. In report [2], an SDK's startup import of the http module added init time on a newer JavaScript runtime.

What are the symptoms of cold start latency regression?

The first invoke of a minimal serverless function was 5 to 10 percent slower on the new runtime version [1]. Function init duration rose from about 204 ms to about 255 ms with the SDK on the newer JavaScript runtime [2].

Why did tests not catch cold start latency regression?

The runtime maintainers' startup measurements used container settings in which the new version was faster [1]. The low-CPU serverless cold start only showed the slowdown when measured directly.

Which test class catches cold start latency regression, and what does it miss?

Load and performance tests and canary checks are listed, both with the by-design label. A performance test misses the regression unless it reproduces the low-CPU serverless cold start [1].

Which public reports document cold start latency regression?

Two primary reports from two organisations document it. They were published on 2025-03-02 [2] and 2025-10-01 [1].

What should be checked before a release to prevent cold start latency regression?

PR-01 measures P90 cold start duration over 100 invokes on the new and current runtime versions. PR-02 runs that benchmark at the lowest memory and CPU setting used in production.

Release Incident Playbook (2026). Cold start latency regression: symptoms, cause, and the test that catches it, content version 1.0.0. https://releasefailures.com/incidents/cold-start-latency-regression/. Accessed 2026-10-21.