Last time I tried (maybe 1-2 years ago), there were some errors when building the doc with napoleon and it was not immediately obvious how to fix them. Googling at the time, it was not immediately clear what the differences were between napoleon and numpy-doc. There may also be small differences in the doc output that are hard to spot given the size of the scikit-learn documentation.

I would be in favour of waiting for 0.20 before merging a PR that switches to napoleon, so that quirks can be spotted in the dev doc until the 0.21 release.

numpydoc has seen some fixes by @jnothman and @amueller semi-recently (maybe a few months ago) so maybe they have some informed opinion.

I'm also not clear on the differences between the two feature-wise (their documentations are not too explicit about this point), I have used both in the past.

However, another point to consider that is that numpydoc had extensive contributions by Joel (and other scikit-learn developers) while, unless I am mistaken no one contributed to sphinx.ext.napoleon too much.

So this would mean replacing a working and well known dependency, by another one which potentially will require more maintenance work. The end result is that new contributors will not need to install numpydocs when they run tests or build docs locally, but is that worth the effort?

What is the original reason to bring this up? Only that napoleon is included in sphinx? Or that we think it actually works better, or is better maintained?

Based on how other project have managed similar issues (matplotlib#5743, astropy-sphinx#5) and the effort put in #21350, I believe this one can be closed.

Sounds good with me. I'm closing it.