Top-posting to recap all the interesting answers and answer my initial mail.
The overall feeling I get is that even with the changes that may be needed to satisfy the new resolver, we should be fine to apply these to stable branches:
* lower-constraints was discussed a lot, this is where largest changes were spotted but they are OK given the current use/effectiveness of these jobs (or maybe even dropped soon)
* linters can be extracted from test-requirements, to limit linters version bumps. I had quickly tried that for the neutron fix and it had failed in some other job, but I will take another look in a separate patch. Then if needed this change can be squashed with pip requirements fixes in stable branches.
* For some recent branches (victoria for example), style fixes are small so this can be just cherry-picked from master to have a working branch
* Other requirements bumps should be OK as they actually indicate the proper needed versions now
* If we ever hit a change (old third-pary dependency) that cannot be fixed without going over upper-constraints, then we may have to cap pip. Hopefully, this will not be hit.
If all of this sounds good, then I guess it will be time to play whack-a-stable-mole
On 2020-12-13 14:39:58 +0100 (+0100), Luigi Toscano wrote:
> On Saturday, 12 December 2020 00:12:36 CET Jeremy Stanley wrote:
> > On 2020-12-11 20:38:30 +0000 (+0000), Sorin Sbarnea wrote:
> > [...]
> > > Regarding decoupling linting from test-requirements: yes! This was
> > > already done by some when conflicts appeared. For old branches I
> > > personally do not care much even if maintainers decide to disable
> > > linting, their main benefit is on main branches.
> > [...]
> >
> > To be honest, if I had my way, test-requirements.txt files would die
> > in a fire. Sure it's a little more work to be specific about the
> > individual requirements for each of your testenvs in tox.ini, but
> > the payoff is that people aren't needlessly installing bandit when
> > they run flake8 (for example). The thing we got into the PTI about
> > using a separate doc/requirements.txt is a nice compromise in that
> > direction, at least.
>
> Wouldn't this mean tracking requirements into two different kind
> of places:the main requirements.txt file, which is still going to
> be needed even for tests, and the tox environment definitions?
Technically we already do. The requirements.txt file contains actual
runtime Python dependencies of the software (technically
setup_requires in Setuptools parlance). Then we have this vague
test-requirements.txt file which installs everything under the sun
a test might want, including the kitchen sink. Tox doesn't reuse one
virtualenv for multiple testenv definitions, it creates a separate
one for each, so for example...
In the nova repo, if you `tox -e bandit` or `tox -e pep8` it's going
to install coverage, psycopg2, PyMySQL, requests,
python-barbicanclient, python-ironicclient, and a whole host of
other stuff, including the entire transitive dependency set for
everything in there, rather than just the one tool it needs to run.
I can't even run the pep8 testenv locally because to do that I
apparently need a Python package named zVMCloudConnector which wants
root access to create files like
/lib/systemd/system/sdkserver.service and
/etc/sudoers.d/sudoers-zvmsdk and /var/lib/zvmsdk/* and
/etc/zvmsdk/* in my system. WHAT?!? Do nova's developers actually
ever run any of this themselves?
Okay, so that one's actually in requirements.txt (might be a good
candidate for a separate extras in the setup.cfg instead), but
seriously, it's trying to install 182 packages (present count on
master) just to do a "quick" style check, and the resulting .tox
created from that is 319MB in size. How is that in any way sane? If
I tweak the testenv:pep8 definition in tox.ini to set
deps=flake8,hacking,mypy and and usedevelop=False, and set
skipsdist=True in the general tox section, it installs a total of 9
packages for a 36MB .tox directory. It's an extreme example, sure,
but remember this is also happening in CI for each patch uploaded,
and this setup cost is incurred every time in that context.
Thanks for the hint btw, I'll apply it to our repos.
I will have to check that too, making these jobs lighter for CI is always nice!
This is already solved in a few places in the nova repo, in
different ways. One is the docs testenv, which installs
doc/requirements.txt (currently 10 mostly Sphinx-related entries)
instead of combining all that into test-requirements.txt too.
Another is the osprofiler extra in setup.cfg allowing you to `pip
install nova[osprofiler]` to get that specific dependency. Yet still
another is the bindep testenv, which explicitly declares deps=bindep
and so installs absolutely nothing else (save bindep's own
dependencies)... or, well, it would except skipsdist got set to
False by https://review.openstack.org/622972 making that testenv
effectively pointless because now `tox -e bindep` has to install
nova before it can tell you what packages you're missing to be able
to install nova. *sigh*
So anyway, there's a lot of opportunity for improvement, and that's
just in nova, I'm sure there are similar situations throughout many
of our projects. Using a test-requirements.txt file as a dumping
ground for every last package any tox testenv could want may be
convenient for tracking things, but it's far from convenient to
actually use. The main thing we risk losing is that the
requirements-check job currently reports whether entries in
test-requirements.txt are compatible with the global
upper-constraints.txt in openstack/requirements, so extending that
to check dependencies declared in tox.ini or in package extras or
additional external requirements lists would be needed if we wanted
to preserve that capability.
--
Jeremy Stanley
--
Red Hat GmbH,
https://de.redhat.com/ , Registered seat: Grasbrunn,
Commercial register: Amtsgericht Muenchen, HRB 153243,
Managing Directors: Charles Cachera, Brian Klemm, Laurie Krebs, Michael O'Neill
--