Flaky test in HNSW index unit test

[git-hulk]

Search before asking

• I had searched in the issues and found no similar issues.

Version

unstable

Minimal reproduce step

See https://github.com/apache/kvrocks/actions/runs/10971334789/job/30466698610?pr=2538#step:15:173

What did you expect to see?

CI passed

What did you see instead?

CI failure

Anything Else?

No response

Are you willing to submit a PR?

• I'm willing to submit a PR!

[LindaSummer]

Hi @git-hulk ,

I want to follow up this issue. 😊
Can it be assigned to me?

Best Regards,
Edward

[git-hulk]

@LindaSummer Sure, thank you!

[LindaSummer]

Hi @Beihao-Zhou ,

Sorry to bother you.
Please correct me if my analysis is incorrect. 😊

The GitHub Action's log is Layer: 4 Frequency: 458 Expected: 411.52266242457773.
If I understand correctly, this expected value should be $(3{^-4}-3{^-5})*50000=411.5226337448559$.

Since we use a random seed from random_device.

kvrocks/src/search/hnsw_indexer.cc

Lines 180 to 181 in f2bc224

	std::random_device rand_dev;
	generator = std::mt19937(rand_dev());

We can't reproduce the same generation sequence and in my knowledge the generation result may not be same as expected.

Could I use a fixed rand seed for unit test startup initialization to solve this issue?

I think it will improve the reproducing ability for test cases.

Best Regards,
Edward

[Beihao-Zhou]

Thanks for looking into the issue! @LindaSummer
Yes sure, I think you can change the interface a bit to pass in the fixed seed. For the current unit tests, I just set a relatively low threshold so that in majority times it can pass even if it's randomized.
Thanks again for fixing it!