If we want to have to have reproducible examples, fixing the random seed should be done in the examples, not in our documentation building system.

(That is also unlikely to work with sphinx-gallery).

I'm not sure I agree about putting it in the examples themselves. IMO, the seed isn't for the benefit of those using them, so much as our ability to use them to smoke out changes. You'd be adding more lines of boilerplate for no educational benefit.

I'm not aiming for reproducible examples, but for reproducible documentation. It means if you change some small documentation thing, only that thing will change.

In a way, I think it's more "fun" (or let's say, interesting) to have the examples be truly random, and it detracts from them by setting a seed.

Couldn't a quick git grep tell us if the examples use random? And then change those over to np.random just to keep this cleaner.

There's only one example importing standard library's random and it's not using it (I just put up a PR removing it in #7385).

I like the idea of reproducible plots in our docs.

I'm not sure what the significance of 20021210 so if it's all the same, and we don't want to just used 0 as the seed, I suggest 19680801 - John Hunter's birthday.

@NelleV mentioned that this won't affect sphinx-gallery when we switch to that, but perhaps it'd be worth adding the plot_pre_code feature of plot_directive we're utilizing here. I added an issue for this in sphinx-gallery: sphinx-gallery/sphinx-gallery#158

On the other hand - there are only 17 files that use np.random, and of those, most are rst docs, that already set the seed, leaving only 7 .py files:

$ git grep  np\.random | cut -f 1 -d\: | uniq |grep \.py 
pyplots/align_ylabels.py
pyplots/boxplot_demo.py
pyplots/compound_path_demo.py
pyplots/dollar_ticks.py
pyplots/fig_axes_labels_simple.py
pyplots/pyplot_scales.py
pyplots/pyplot_text.py

of those, pyplots/boxplot_demo.py already sets a seed, so the alternative to writing the seed into the examples isn't that onerous either.

@ivanov then converting those docs to using set seeds sounds like a great little project for new contributors or @NelleV's machine shop if they still need tasks. Unless you just want to throw that into #7385...

Why are you only looking at pyplots? All of the examples are included in the docs.

I could have sworn it was >100, but I must have forgotten to exclude the seeded ones. There are 80 examples without seeds:

$ grep -L 'RandomState' $(grep -L seed $(grep -IRl 'np.random' mpl_examples/)) | wc -l
80

plus the 6 in pyplots that @ivanov noted above and I forgot to count.

Oh, and 20021210 was the first date I could pull out of the changelog, but I have no real tie to it.

Why are you only looking at pyplots? All of the examples are included in the docs.

I used git grep - which doesn't follow symlinks. I did not realize that mpl_examples was symlinked to ../examples and therefore none of the ~100 examples that use np.random in there showed up in my original search. So I retract my previous suggestion of just adding a seed to all of those.

I still think this could be a good machine-shop or new contributor task as part of the general docs overhaul...but that probably shouldn't be a blocker for this/this code can be taken out if that ever happens.

I can volunteer to make the examples reproducible.

Closing for #7386.