Closed hombit closed 6 months ago
My opinion: when someone sees different results in recurring sessions that reveals an error in code. When we fix random seeds some bugs will be covered and eventually they will have a longer lifespan.
I would not say that I am 100% agree with @hombit , but we can explicitly use now()
as a random seed for the fuzzy testing purposes to uncover this bugs.
@matwey @balodja the issue is driven by our own experience in the SNAD team, when few years ago we found that we cannot reproduce some results because of missed random state
I'm happy to have None, or now(), as a random state in tests. My point is that we should try to make life of the users with real data and real expert time spent on analysis more important than test writing experience.
We focus on reproducibility, especially for active learning sessions. I believe the better default is having constant random seeds in all the places