I should probably add that the CircleCI builds are failing with an error that I saw myself when building, and when I said the patch compiles with sphinx I meant with the -W switch omitted. Considering how the same warnings are present in the original v2.2.3-doc branch I suspect the source of the test failures is not in my changes.

For this one I don't know what the best route of action is. I guess it makes sense to have it in the "what's new in 2.2.3" of the 2.2.3 docs, but of course in the "Previous what's new" of any subsequent release as well. Would that require two different PRs to be created, one against master and one against the 2.2.3 branch?

Aren't the "Previously what's new" items just links to the generated what's news of earlier version? Is there a "What's new in 2.2.3" other than this one?

Nevermind, you're right of course, found it. Since 2.2.3 is LTS for python 2 users we should edit both I believe.

Submitted #12184 for master because it's easier to ask for forgiveness.

The circle failures are understood, we have to backport a bunch of commits to this branch to cope with the sphinx 1.8 changes.

@adeak the failures should be sorted now, would you be able to rebase this on to the current master branch?

@dstansby I'm afraid I'm confused. Master should correspond to 3.0, so its doc/users/whats_new.rst is incompatible with the 2.2 What's New that I'm trying to fix in this backport. But the v2.2.3-doc branch where I thought this would belong hasn't seen a new commit since I opened this PR. What am I missing?

I would have though it's enough to rerun the tests. I tried to trigger this by closing and reopening, but it won't work the circleci tests.

@ImportanceOfBeingErnest I touched the PR with a no-op commit amend; this seems to have triggered all the tests.

No luck; though I'm not terribly surprised considering that the code base hasn't changed since I opened the PR. After Ernest's comment I was hoping there was some configuration change or something in the background that could have fixed the issue regardless.

Ha, it's even worse, now travis fails as well for some unrelated reason. I think one has to wait until this gets fixed - those tests are beyond my paygrade.

I would have though it's enough to rerun the tests. I tried to trigger this by closing and reopening, but it won't work the circleci tests.

Unfortunately, CircleCI builds the head of the PR, not the merge of the PR and the target, like the other CIs do. That being said, there were no backports to the v2.2.3-doc branch, so it's unlikely for anything to work anyway.

This should work now if rebased, for doc builds only. Travis fixes are still in progress.

Thank you, all, indeed Circle works now and Travis breaks still.

Well, that fix has not come back to this branch anyway.