Flaky test: test_distributions_random.py::TestNestedRandom::test_TruncatedNormal

[michaelosthege]

A wild, flaky test has appeared.

On master it's currently failing here: https://github.com/pymc-devs/pymc3/runs/1562295314#step:8:1434

I already observed the flaky tests in a recent PR but it went away after re-running the CI.

[Sayam753]

Same flaky test appeared in https://github.com/pymc-devs/pymc3/runs/1638797078

[twiecki]

Probably just needs a seed.

[Sayam753]

Another one https://github.com/pymc-devs/pymc3/pull/4416/checks?check_run_id=1714255767

[michaelosthege]

Amazing how often they appear. And always in the same test, but sometimes with different parametrizations?!

But I just added the "wontfix" label, because all the random stuff will be replaced as we switch to RandomVariable. Let's hope it goes away then.

[Sayam753]

Sounds great 🤩. Is there a doc listing all the features that RandomVariable Op will add?

[michaelosthege]

Sounds great 🤩. Is there a doc listing all the features that RandomVariable Op will add?

Not for the RandomVariable Op that is already implemented in Theano.
However there is Brandon's blogpost from three years ago when he birthed the idea: https://brandonwillard.github.io/random-variables-in-theano.html#orged59f09

My understanding is that essentially the .eval() of the TensorVariable will result in drawing random samples - also on dependent variables that are created from a RandomVariable.
This way we don't need to walk up the graph in PyMC3 as we currently do in draw_samples and the like.

[ricardoV94]

Test seems to be stable for a while now