Introduce pm.config or rcParams for global & local config settings

[michaelosthege]

Let me begin with saying that global variables are evil and nobody should be using them.

For things like #4372 it would be really useful to have a place where to make global config changes.
Matplotlib does it and ArviZ too.
Aesara also has a quite powerful config implementation that I refactored quite a bit a few months back.
Or maybe we don't even need/want a place for global configuration settings?

Anyways it would be great to discuss and figure out what we want!

Update 2022-03-30 + ToDo

We have a general consensus that we want to do this. (I'm updating the title+labels in a second.)

Whoever wants to work on this should

1. Read this thread
2. Specifically investigate the options here
3. Decide which route to go for
4. Open a PR

[ricardoV94]

What kind of things would go in here?

[michaelosthege]

• Switching the return_inferencedata default value
• Turning the automatic addition of SpecifyShape when full-length shape is passed on/off
• Default values for compute_convergence_checks or discard_tuned_samples maybe?

[OriolAbril]

• default init in sample, and anything else where we have a "you can't trust the default to stay the same"
• keep_size ? for people who wants to live in the past and get flattened samples in sample_posterior_predictive after Set keep_size=True in v4 #4647
• going forward not sure if it should be here or in Aesara, default sampling backend, jax, numba, python samples if still around

[twiecki]

I like it, I often use a default-dict for the sample() kwargs.

[michaelosthege]

Sounds like we are in favor of global config settings.
So how should we implement it?

Option 1: rcParams

➕ Users know it from matplotlib and arviz
➕ Can change options with context managers (I didn't know that - thanks @OriolAbril !)
➖ Mix of two different config schemes (Aesara vs. PyMC3)

Option 2: aesara.configparser

➕ Users know it from theano/aesara
➕ Can change options with context managers
➕ Can work with CLI params/environment variables
➖ Mix of two different config schemes (Aesara+PyMC3 vs. ArviZ)

Feel free to edit this comment to add/edit the ➕/➖ items.

[OriolAbril]

I'd go with whatever is easier to add and maintain. ArviZ/matplotlib version has no dependencies because it has the reader and rcparams object as part of the codebase.

Also not sure what you mean by always global in the matplotlib/arviz one, both can be configured temporarly with context managers.

[michaelosthege]

After looking at the size of the ArviZ rcParams implementation 😮 I'd actually recommend the Aesara configparser.
It comes down to copyediting the following:

• A_create_default_config function in pymc3/configdefaults.py
• Creating a config instance in pymc3/configdefaults.py and importing it to the package level: pm.config
• Adding default settings with pm.config.add() as seen in aesara/configdefaults.py

[ferrine]

I like the idea. There should also be a way to make temporary changes in a context manager like done in matplotlib