Fix a flaky test

A flaky test is a real defect, in the test or in the product. Retrying hides it and teaches everyone to ignore red builds. The aim is to reproduce it on demand, fix the cause, and show that the same conditions no longer break it.

When to use

A test fails intermittently, or passes when rerun without changes.

Steps

1. Collect the failures. Save the exact message, OS, job and commit. Note what the failing runs have in common.

2. Reproduce it. Try these until it fails:

   • Run the test many times in a loop.
   • Run it with other tests, in a different order, or in parallel.
   • Add CPU load, or limit the CPU.
   • Slow down the code it waits for with a temporary delay.

   Slowing the code is often the quickest way to prove a timing bug.

3. Classify the cause.

   Type	Signs	Usual fix
   Fixed wait too short	Fails on slow hosts	Wait for the outcome
   Shared state	Depends on test order	Give each test its own state
   Real clock	Fails near midnight or a boundary	Inject the clock
   Unordered results	Order changes between runs	Sort or compare as a set
   Port, file or process conflict	Fails in parallel	Use unique resources
   Real race in the product	Wrong result under load	Fix the product
   External dependency	Fails with the network	Stub the boundary

4. Fix the cause. Replace a fixed wait with a check on the actual outcome. A deadline only catches a hang, so set it longer than the time the product allows. Make sure a timeout prints enough output to explain itself.
5. Check the assertions are unchanged. The fix must not make the test accept behaviour it used to reject.
6. Prove the fix. Repeat the conditions from step 2. The old code must fail and the new code must pass under the same stress.
7. Finish. Follow verify-and-report.

Stop and ask when

• You cannot reproduce it after trying everything in step 2. Report what you tried, and add diagnostics instead of guessing.
• The cause is a real product race that needs a design change.

Common mistakes

Example

Failure: replaced=false after 5 seconds, on Windows runners only.

Weak fix: raise the wait from 5 to 15 seconds.

Strong fix:

Reproduce: slowed the old process's shutdown to 6 seconds, and the test failed
with the same message as CI.
Cause: the test stopped after a fixed 5 seconds, while the handoff can take longer.
Fix: wait until the handoff finishes, with a 60-second bound for hangs.
Proof: with the same 6-second delay, the old test fails and the new one passes.
Assertions: unchanged.

Report

Failure: <message, where it occurred>
Reproduced by: <command or condition>
Cause: <type and explanation>
Fix: <what changed>
Proof: <old fails, new passes, under the same condition>
Assertions changed: <none, or why>