On Thu, Feb 22, 2018 at 03:50:38PM +0100, Pierre-Yves Chibon wrote:
On Wed, Feb 14, 2018 at 05:28:20PM +0100, Petr Šplíchal wrote:
Hi!
During the last days there have been concerns raised regarding
what is an appropriate content for the tests namespace. [1] My
original idea was to enable sharing tests even across branches of
the same component, not only for tests to be used by completely
different packages. The initial examples might have been a bit
misleading in this respect. One of the main points still holds:
* Tests might follow a different branching pattern than the
dist-git repo, leading to code duplication
From the feedback from developers I feel they always keep on mind
and care a lot about the maintenance costs. So it perfectly makes
sense to me if they want to keep and maintain tests in a separate
repo instead of merging/cherry-picking between dist-git branches.
When, for a particular package, it is the most efficient way to
maintain tests in a separate repo why should we discourage from
this approach? There are packages where it makes more sense to
store test code directly in dist-git. And it is still an option.
But why should we enforce this for all?
My worries are basically that this mechanism is hiding the tests from the
package maintainers. It splits the concerns between people maintaining the
artifacts and people maintaining the tests, which is exactly what the original
plan to bring CI was *not* about.
The idea has always been to bring the tests where the code lives, so that both
can be worked on as one, to make tests be a concern of the package maintainer
very much like updating the artifact (rpms, image..) is, while getting support
from QE for them (the tests).
In addition, this is what I fear most:
The tests for the package are stored elsewhere, we're not sure where (the tests
namespace, a random git repo on the internet...), the packager update package
and the tests fail.
What do you think will happen?
a) the packager will look for $random_place_of_the_internet where tests are and
spend time trying to fix them.
b) the packager will turn off the tests because they get in the way
If the packager wants to turn off the tests, they will have to go through
dist-git to do it, they may very well end up turning the tests off anyway, but
if the tests are right there, they may as well have a quick look at them to see
if they can fix them quickly before deciding.
In addition, if the packager turn the tests off and the people maintaining the
tests do not realize that, they will end up spending time maintaining
$somewhere_else tests that aren't being used.
Having them interact directly with the dist-git repo will increase the chances
that they see/realize when something is turned off.
If that means we have less tests in dist-git because the maintainers do not want
them, I'd say so be it.
In the long term this is their loss, they are the ones who will get the bug
reports and will have to deal with them, they are the ones who will have to dive
into old code rather than going back into something that is still fresh in their
mind.
As long as, it is clear that they have been approached and that it is their
choice to not merge pull-requests adding tests, I think the people offering to
help should not be the ones blamed.
pingou, I share your thinking in general, but I think your concern is
overstated. Petr's original e-mail suggests that the new separate
namespace should be used *in* *preference* to the in-dist-git, and your
reply concentrates on that. But I think that we want to have both
possibilities, because both will work best for different cases, and it's
just the emphasis that needs to change.
I agree that having tests in a separate namespace is not any simpler
or more manageable for most packages. Thus, I think that *by* *default*
we should put tests in dist-git. IMHO this will apply to 90% of all
tests and packages.
OTOH, there will be cases where it'll be useful to split the tests out
into a separate area. The documentation for tests/ has the example of
shared tests for bash/zhs/dash. I'm sure that there will be other
cases, like testing of stacks consisting of multiple packages
apache+mysql+php or glibc+gcc+binutils, etc, where it'll be beneficial
to keep the test separate. The management of permissions in this case
will be _much_ more complicated, because to meaningfully update the
tests sometimes it'll be necessary to change the test repo and _each_
of the packages. But having them separate is better than randomly
sticking the shared tests in dist-git for one of the involved packages.
Nevertheless, because of the overhead of more repos and additional
permissions, IMHO this should never be advertised as the default
approach, but as an "advanced" possibility.
