Re: Exploring Sphinx, autodoc, apidoc, and coverage tools for python/qemu

[John Snow <jsnow@redhat.com>]

On 7/24/19 5:06 PM, John Snow wrote:
> Has anyone on this list experimented with these tools?
> 
> I was hoping to use them to document things like the python/machine.py
> and python/qmp.py modules to help demonstrate some of our internal
> tooling API (for test writers, GSoC/Outreachy interns, folks who want to
> script QEMU at a level between writing a CLI driver and using libvirt.)
> 
> What follows below is my process trying to enable this and some of the
> problems I'm still stuck with, summarized below at the end of this more
> exploratory text.
> 
> 
> Enabling autodoc:
> 
> First, it appears as if enabling the "sphinx-autodoc" tool is not
> sufficient for actually generating anything at all when you invoke the
> sphinx-generated "make html" target. It just enables understanding
> certain directives.
> 
> So apparently you need to generate module "stubs" using sphinx-autodoc.
> Sphinx uses the sphinx-autodoc extension to understand how to consume
> the directives in these stubs.
> 
> That strikes me as odd, because these stubs might need to be changed
> frequently as code comes and goes; it seems strange that it isn't
> integrated at the top level. (Do I have the wrong idea on how these
> tools should be used?)
> 
> So you need to run:
>> sphinx-apidoc --separate --module-first -o docs/ python/qemu/
> 
> which generates stubs to docs:
> 
> Creating file docs/qemu.machine.rst.
> Creating file docs/qemu.qmp.rst.
> Creating file docs/qemu.qtest.rst.
> Creating file docs/qemu.rst.
> Creating file docs/modules.rst.
> 
> And then you can edit e.g. the top-level index.rst TOC in docs/index.rst
> to look like this:
> 
> ```
> .. toctree::
>    :maxdepth: 2
>    :caption: Contents:
> 
>    interop/index
>    devel/index
>    specs/index
>    modules
> ```
> 
> And then finally generating the build; manually removing the -W option
> from the Makefile: there are a lot of warnings in here.
> 
>> sphinx-build -n -b html -D version=4.0.92 -D release="4.0.92
> (v4.1.0-rc2-34-g160802eb07-dirty)" -d .doctrees/
> /home/bos/jhuston/src/qemu/docs/ docs/
> 
> Great! that will generate output to docs/index.html which indeed shows
> APIdoc comments generated from our Python files. Good.
> 
> However, where this gets a little strange is if you look at the
> generated stubs. For instance, qemu.machine.rst looks like this:
> 
> ```
> .. automodule:: qemu.machine
>     :members:
>     :undoc-members:
>     :show-inheritance:
> ```
> 
> :undoc-members: says that we want to "document" any members that don't
> have a matching apidoc comment by generating a stub.
> 
> Oops, but the presence of that stub will cause the sphinx coverage tool
> to happily report 100% coverage.
> 
> Further oops, pylint doesn't understand apidoc comments and can't be
> used as the linter in this case, either.
> 
> You can edit the stubs to remove these directives, but these stubs are
> generated -- and it doesn't appear like there's a command line option to
> change this behavior. ...Hmm.
> 

Update: there is indeed a way to change this behavior, through
environment variables.

https://www.sphinx-doc.org/en/master/man/sphinx-apidoc.html

SPHINX_APIDOC_OPTIONS
A comma-separated list of option to append to generated automodule
directives. Defaults to members,undoc-members,show-inheritance.

You just have to set it and remove 'undoc-members'.

> And either way, the coverage tool only generates a report and not
> something with an error code that I could use to gate the build. Same
> goes for the general build: if I remove the :undoc-members: parameter,
> there's nothing in the autodoc module that appears to throw warnings
> when it encounters undocumented parameters or members.
> 

This is still a problem, though. Nothing gates on undocumented members
at all.

> That seems disappointing, because it's hard to keep docstrings up to
> date unless they are checked conclusively at build time.
> 
> 
> Conclusions:
> 
> - the autodoc documentation page doesn't seem to document examples of
> how you're expected to write meaningful docstrings for the tool to extract.
> 
> - autodoc fools the coverage tool into reporting 100% coverage.
> 
> - autodoc can be configured to omit non-documented members to allow the
> coverage tool to work, but the configuration is auto-generated and
> defaults to always generating documentation for these entities.
> 

apidoc (the stub generator for autodoc) can indeed be configured to omit
non-documented members through an environment variable now, so this
point is not true.

> - coverage tool doesn't appear like it can be used for gating the build
> natively for missing python docs; it only generates a report.
> 

Still seems true, though there are other tools that can gate on such
things. Specifically we can use pylint and require that docstrings are
present.

Notably, pylint only checks that there is a docstring at all, and
doesn't try to perform any semantic validation of what's in it.

To my knowledge, there are no standalone tools that do such work -- but
the pycharm built-in linter appears to have semantic support for certain
docstring formats.

> - Even if we script to block on a non-empty report, the coverage tool
> only works at the function/class level and does not understand the
> concept of missing parameter or return value tags.
> 

Still a problem based on the above; there is just no agreed upon
semantic format in the python world for documenting parameters.

> - It would seem that it would be the Autodoc module's job to be
> responsible for understanding incomplete documentation, but doesn't
> appear to. The :param name: syntax is just a ReST "field list" and isn't
> parsed semantically by autodoc, sadly.
> 

Might be something I investigate based on some discussions I've had in
the Python community. Could be useful to have a canonical docstring
format that sphinx is able to give warnings against.

> 
> It looks to me, at a glance, that there's nothing in Sphinx that knows
> how to look for and warn about undocumented parameters, exception types,
> return values, etc. Hopefully I've missed something and it is possible.
> 
> --js
>