issue_comments

looking at this once again, there are a lot of projects that don't install (eg aiohttp (build), but I'm not sure how this is set up or which parts of their configuration comes from somewhere other than their .readthedocs.yml).

Also, it seems we are using the old version of readthedocs.yml.

And finally, it seems strange to me that the scipy19-docs fails but master doesn't, which means that master does something different?

As I don't have access to readthedocs I can't really experiment on these suggestions (making this informed guessing at best), so I would like to leave the tracking down and fixing to someone else.

I'm currently not planning to because I will probably be offline for the next few days

Edit: the only workaround I can imagine is building the docs without installing (which is mentioned in the docs), but I neither know why the current configuration is the way it is or how building the docs without installing works with RTD

it seems all of these problems boil down to that single command. What gets executed is (essentially) bash $ python -m pip install --upgrade --upgrade-strategy eager . which upgrades pandas to 0.25.1, but leaves dask as is because it is an optional dependency. This, however, defeats the purpose of version pinning in the conda file (or a requirements file).

I can't find an option to tell readthedocs to install without the eager upgrade (normal upgrade is fine). Maybe ask support? The addition of the upgrade strategy happened in readthedocs/readthedocs.org#5635

After some experimentation, it seems this is only the case if setup.py install is used, not with pip install: for me, at least, in the commands from https://github.com/pydata/xarray/pull/3199#issuecomment-534342306, replacing bash $ python setup.py install --force with bash $ python -m pip install . makes this work, even without the merge commit.

The docs on the settings mention pip_install, maybe use that?

The problem is still the same: while the output of the environment printed by sphinx lists the new version as xarray 0.13.0+22.gaeb15b56 pypi_0 pypi which in theory should be used, but building the docs prints this: xarray: 0.13.0, /home/docs/checkouts/readthedocs.org/user_builds/xray/conda/scipy19-docs/lib/python3.7/site-packages/xarray/__init__.py which does not have the version of open_dataset that allows non-nc files. Honestly, I'm no expert on imports and version management, so I don't know what the best fix for this is.

Something that comes to mind is waiting for a new release with the new version of open_dataset, but that is not an immediate fix (and only a workaround at best).

err, well, this is somewhat confusing: I'm not sure what pulls it, but it seems something depends on xarray (installed in the creation of the environment). The version that is conda-installed is 0.13.0, so as master does not have the new version of open_dataset that does not enforce ".nc" files, the 404 happens. In short, the version from setup.py install gets shadowed, I guess because the version is 0.12.3? In that case, merging master into scipy19-docs might fix this?

(roughly) following the instructions in the build log, I can reproduce it raising a 404: bash $ conda env create --quiet --name scipy19-docs --file doc/environment.yml $ conda install --yes --quiet --name scipy19-docs mock pillow sphinx sphinx_rtd_theme (scipy19-docs)$ python -m pip install -U recommonmark readthedocs-sphinx-ext (scipy19-docs)$ python setup.py install --force (scipy19-docs)$ cd doc; make html I didn't check yet, but maybe the second conda install should not happen, as all those get installed on env creation?

that seems to have worked, the failures should be unrelated.

unless something else is wrong, either modifying the md5 file (remove the newline) or adding strip() to f.read() in tutorial.open_dataset should fix the md5 check.

I'm not sure if this helps you, but I noticed two issues in xr.tutorial.open_dataset: first, it requires the file extension to be ".nc" (resulting in a 404 when fetching the file. The version from the scipy19-docs branch does not have that issue), and second, there is a newline in the md5 file. Modifying remotemd5 = f.read() to remotemd5 = f.read().strip() should do the trick.