New approach to handling randomness in pyMOR

[sdrave]

This PR implements a new approach to handling a global random state in pyMOR. Most notable changes compared to the previous approach:

• The newer np.random.Generator class is used instead of np.random.RandomState.
• seed and random_state arguments are removed from all function signatures. Instead to execute some function with a prescribed random state, a new global random generator can be locally installed by used the RNG object returned by new_rng as a context manager:

  with new_rng(seed):
      the_random_function(...)

• In concurrent code, the spawn_rng wrapper can be used to create child RNGs to ensure deterministic and uncorrelated behavior. Not calling spawn_rng where needed will be detected and cause warnings to be emitted.

[sdrave]

@pymor/main, @artpelling this should be roughly what we have discussed at the sprint last week. Please take a look if it makes sense to you. (I did not look at the notebooks/tests yet, and documentation is still missing.)

[sdrave]

@renefritze, @pmli, I have reworked how the random state is set/initialized: init_rng/set_rng are now replaced by a new_rng function, which returns a pyMOR RNG class, which directly derives from numpy.random.Generator. Besides being directly usable as a random generator, RNG has install/uninstall methods, which sets it as pyMOR's global rng. Further it can be used as a context manager to automatically install/uninstall it. In particular, you can now create a single new RNG object and repeatedly set it inside loops and so on.

[codecov]

Codecov Report

Merging #1736 (8a34c7f) into main (8229a85) will decrease coverage by 0.00%.
The diff coverage is 92.26%.

Additional details and impacted files

Flag	Coverage Δ
github_actions	75.19% <85.08%> (+0.03%)	⬆️
gitlab_ci	86.98% <92.26%> (-0.01%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
src/pymordemos/parabolic_mor.py	97.08% <ø> (ø)
src/pymordemos/thermalblock.py	95.23% <ø> (ø)
src/pymor/algorithms/lradi.py	87.50% <50.00%> (-0.53%)	⬇️
src/pymor/algorithms/lrradi.py	87.27% <50.00%> (-0.54%)	⬇️
src/pymor/reductors/neural_network.py	88.72% <75.00%> (+0.21%)	⬆️
src/pymor/algorithms/samdp.py	88.96% <80.00%> (+0.03%)	⬆️
src/pymor/tools/random.py	86.88% <86.44%> (-13.12%)	⬇️
src/pymor/algorithms/eigs.py	94.40% <100.00%> (+0.09%)	⬆️
src/pymor/algorithms/hapod.py	90.10% <100.00%> (+0.05%)	⬆️
src/pymor/basic.py	100.00% <100.00%> (ø)
... and 21 more

[artpelling]

Looks good to me (I might have found some commas). I tested it locally with my randomized LinA stuff and it works like a charm!

[artpelling]

Closes #1564.

[sdrave]

Ok, so after rebasing CI segfaults on gitlab. @renefritze, any clue why?

[sdrave]

Seems like some/all other current PRs are affected as well.

[renefritze]

Ok, so after rebasing CI segfaults on gitlab. @renefritze, any clue why?

None.

[sdrave]

This seems to be a related but different issue with codecove, @renefritze?

[renefritze]

This seems to be a related but different issue with codecove, @renefritze?

Yeah, no idea what's even failing there. Turned both total and patch coverage targets to 0 now.