Open Stack

Unit tests I've seen in a number of OpenStack projects have been significantly less than perfect and I think we all know this (I hope).
 
We need to make a change ASAP because it is only going to get harder. I don't think just saying it's too hard and continuing to do as we currently do is going to do us any good in the long run. We must create a good system of enforcement when it comes to writing good unittests and having functional test coverage.
 
I'm not sure what the most effective course of action here is but I do know that we need to get all of the reviewers on the same page and enforcing some guidelines. In the past we have had resources like  [http://wiki.openstack.org/SmallTestingGuide] http://wiki.openstack.org/SmallTestingGuide. Perhaps we need to expand on the definition of good testing habits in OpenStack, come up with a way to promote awareness, enforce guidelines, and fix old and bloated tests.
 
Anyone have any thoughts on some actions we can take? I'd love to talk about this at the summit but hopefully we can make some improvements in the mean time?
 
-Alex
 
-----Original Message-----
From: "John Griffith" <john.griffith at solidfire.com>
Sent: Sunday, February 10, 2013 12:37pm
To: "OpenStack Development Mailing List" <openstack-dev at lists.openstack.org>
Subject: Re: [openstack-dev] unit tests





On Sun, Feb 10, 2013 at 3:03 AM, Huang Zhiteng <[mailto:winston.d at gmail.com] winston.d at gmail.com> wrote:
John, Monty, Mark,

 I think one example is unit tests involving test case against entry
 points which requires installation.  I guess that's John's most
 annoying thing.

 For this special case, I think switching from run_tests.sh to tox (or
 even ditching run_tests.sh) is the right solution.  I'm working on a
 patch for Cinder to make this change
 ([https://review.openstack.org/#/c/21597/] https://review.openstack.org/#/c/21597/).   It's not ready yet as I
 met some difficulty with coverage tests, but as a starter, you can try
 functionality other than coverage tests.



 On Sun, Feb 10, 2013 at 5:23 AM, Monty Taylor <[mailto:mordred at inaugust.com] mordred at inaugust.com> wrote:
 >
 >
 > On 02/09/2013 01:59 PM, Mark McLoughlin wrote:
 >> Hi John,
 >>
 >> On Sat, 2013-02-09 at 12:15 -0700, John Griffith wrote:
 >>> We seem to be going down a path of requiring more and more infrastructure
 >>> and reliance from modules inside of the respective projects code base, and
 >>> lately (the most annoying thing to me) is actually requiring the
 >>> installation of the project and it's services in order for our unit tests
 >>> to run.
 >>
 >> Got an example?
 >
 > Yes. I'd also like an example.
 >
 > FWIW - I agree with the sentiment that our unittests should be simple to
 > run and that cross-project testing should be the purview of integration
 > tests.
 >
 > Monty
 >
 > _______________________________________________
 > OpenStack-dev mailing list
 > [mailto:OpenStack-dev at lists.openstack.org] OpenStack-dev at lists.openstack.org
 > [http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev] http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev



--
 Regards
 Huang Zhiteng



 _______________________________________________
 OpenStack-dev mailing list
[mailto:OpenStack-dev at lists.openstack.org] OpenStack-dev at lists.openstack.org
[http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev] http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
Sorry for dropping the email then going MIA.

There are multiple things I wanted to get some thought around, I'll start with what I believe are the easier ones first:

1. I don't believe that we should require the installation of Cinder to run the Cinder unit tests
 This to me clearly crosses the line in to functional or integration territory, and moving it in to
 a venv (tox, or otherwise) doesn't address my issue with it.  Sorry Winston but right now 21597 does 
 exactly what I don't want.

 Today for Cinder I have the test-requires packages and other required items all installed on my laptop and other machines.  
 I can clone the Cinder repo and run all of the tests using 'run_tests.sh -N' in less than a minute.  Lately unit tests are coming
 online that depend on the egg info being installed as Winston touched on.

2. I haven't figured out what Tox versus Nose buys me in Cinder.
 I understand in Nova and others there may be advantages due to parallelism and some other enhancements, but in my
 case my test cycle time has actually increased by having to install the tox venv when I do a fresh clone, and in the case
 of Nova, I *think* my tests are running a bit faster but really I don't feel that I actually know what's being run and the results.

These first two items are the main things I'm looking for feed-back from folks on, the rest of this email is more around test philosophy, and I don't think I'm interested in trying to come to an agreement or change on things here.  I'd just like to raise some awareness, if anybody is interested in pursuing some of this, maybe we can all get together and chat at the summit.  If people think it's ridiculous that's fine too.


In my mind you have unit tests, functional tests and integration tests.  Unit tests to me aren't designed to find bugs, or test for regressions.  They're specifically to test a unit of code that I've written, and the tests should only depend on that unit of code that I'm intending to test.  

So for example if I write a Cinder driver called 'super_cool_driver', I have a single file of unit tests that tests super_cool_driver methods and NOTHING else.  It has no dependencies on any other files/modules in the Cinder project.  It's only purpose is to make sure that super_cool_driver's behaviours are in fact what I intended and expect.  It also ensures that somebody later doesn't change those behaviours inadvertently.

Testing the interaction with other Cinder modules should be done through functional tests.  This is where devstack exercises would come in to play.  Currently I don't think we put enough emphasis here, when somebody submits a bug fix for something we often say "You need a unit test to catch this sort of thing in the future", maybe that's true in some cases but I'd argue that we should be saying "You need a test in devstack or tempest to catch this in the future".  One of the main reasons unit tests aren't very helpful here IMO is that in most cases a bug results from a refactor or enhancement in the code.  Most of the time when folks make these enhancements they also modify the unit test to behave the way they "expect" it to and think it should.  The result is they just create a unit test that verifies the incorrect behaviour.

Another point that came up; we have "fake" volume objects and such spread out and created in multiple places throughout the Cinder unit tests.  Some folks use the fake class that's set up, and most create a fake in their setup, or on the fly in the test case they're writing.  This creates a bit of a problem if the volume object ever changes (you have to find every fake that's created and modify it, then modify any test that uses it etc etc).  It would be great if at some point in the future we had good fakes for things like volume objects, instance objects etc (sort of along the lines of what Robert mentioned, but I'm not sure I'm thinking of the same scope as he mentioned).

An example of this was recently Sean Dague asked me about test objects for Cinder (Volumes, Snapshots, Types etc), I was kinda surprised but we didn't have an existing set of updated fakes to represent these things.  Seems like it would be smart to have these things (both objects and DB entries) to use throughout the unit tests to provide consistency and make tests a bit more realistic, rather than just rolling your own fake object to fit the test your writing it for.

Thanks,
John
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20130210/426edb10/attachment.html>