BUG: Flaky test `test_truncation_specialized_op`

[michaelosthege]

Describe the issue:

It failed in #6304 and on main: https://github.com/pymc-devs/pymc/actions/runs/3665014447/jobs/6195812078

Reproduceable code example:

pytest -v pymc/tests/distributions/test_truncated.py::test_truncation_specialized_op

Error message:

FAILED pymc/tests/distributions/test_truncated.py::test_truncation_specialized_op[observed] - assert False
 +  where False = isinstance(<pytensor.tensor.elemwise.Elemwise object at 0x7fdf5924e1a0>, TruncatedNormalRV)
 +    where <pytensor.tensor.elemwise.Elemwise object at 0x7fdf5924e1a0> = Elemwise{identity}(AdvancedIncSubtensor{inplace=False,  set_instead_of_inc=True}.0).op
 +      where Elemwise{identity}(AdvancedIncSubtensor{inplace=False,  set_instead_of_inc=True}.0) = xt.owne

PyMC version information:

main

Context for the issue:

No response

[ricardoV94]

I cannot reproduce the failure locally. Can someone else?

[ricardoV94]

Failed again in https://github.com/pymc-devs/pymc/actions/runs/3693633480/jobs/6253850355

It is very strange. Looks like the dispatch may be broken? I can't reproduce it locally

[michaelosthege]

Just a wild speculation, but what if somewhere during graph rewrites there's a non-deterministic order of operations? Could happen by iterating over a set, for example.

[ricardoV94]

No, the part of the test that failed is that the dispatch of the specialized truncation Op Truncated(NormalRV) -> TruncatedNormalRV didn't happen and it returned the fallback implementation.

[larryshamalama]

Just to chip in after some weeks: I also still can't reproduce this bug locally... I found out that it just failed recently again here

[larryshamalama]

@michaelosthege With your recent PR merged, this test seems to now pass...? Did you make any change to the test or the code?

https://github.com/pymc-devs/pymc/actions/runs/4084608130/jobs/7041511557#step:7:34

[ricardoV94]

This is flaky, I wouldn't read to much to it having passed once

[michaelosthege]

AFAIK this test is still flaky - I just got lucky in my PR and merge. Also, my PR was unrelated

[ricardoV94]

Found the reason, that tests sets observed=np.empty(100), which sometimes PyMC converts to a maskedarray, and then we get automatic imputation on the TruncatedRV! It's non deterministic, because it depends on the contents of np.empty

[larryshamalama]

I see the non deterministic aspect of np.empty(100), but why would this inhibit substitution the Op to a TruncatedNormalRV?