Open Stack

On 23 September 2014 03:58, Matthew Booth <mbooth at redhat.com> wrote:
> If you missed the inaugural OpenStack Bootstrapping Hour, it's here:
> http://youtu.be/jCWtLoSEfmw . I think this is a fantastic idea and big
> thanks to Sean, Jay and Dan for doing this. I liked the format, the
> informal style and the content. Unfortunately I missed the live event,
> but I can confirm that watching it after the event worked just fine
> (thanks for reading out live questions for the stream!).
>
> I'd like to make a brief defence of faking, which perhaps predictably in
> a talk about mock took a bit of a bashing.
>
> Firstly, when not to fake. As Jay pointed out, faking adds an element of
> complexity to a test, so if you can achieve what you need to with a
> simple mock then you should. But, as the quote goes, you should "make
> things as simple as possible, but not simpler."
>
> Here are some simple situations where I believe fake is the better solution:
>
> * Mock assertions aren't sufficiently expressive on their own
>
> For example, imagine your code is calling:
>
> def complex_set(key, value)
>
> You want to assert that on completion of your unit, the final value
> assigned to <key> was <value>. This is difficult to capture with mock
> without risking false assertion failures if complex_set sets other keys
> which you aren't interested in, or if <key>'s value is set multiple
> times, but you're only interested in the last one. A little fake
> function which stores the final value assigned to <key> does this simply
> and accurately without adding a great deal of complexity. e.g.
>
> mykey = [None]
> def fake_complex_set(key, value):
>   if key == 'FOO':
>     mykey[0] = value
>
> with mock.patch.object(unit, 'complex_set', side_effect=fake_complex_set):
>   run_test()
> self.assertEquals('expected', mykey[0])
>
> Summary: fake method is a custom mock assertion.
>
> * A simple fake function is simpler than a complex mock dance
>
> For example, you're mocking 2 functions: start_frobincating(key) and
> wait_for_frobnication(key). They can potentially be called overlapping
> with different keys. The desired mock return value of one is dependent
> on arguments passed to the other. This is better mocked with a couple of
> little fake functions and some external state, or you risk introducing
> artificial constraints on the code under test.
>
> Jay pointed out that faking methods creates more opportunities for
> errors. For this reason, in the above cases, you want to keep your fake
> function as simple as possible (but no simpler). However, there's a big
> one: the fake driver!
>
> This may make less sense outside of driver code, although faking the
> image service came up in the talk. Without looking at the detail, that
> doesn't necessarily sound awful to me, depending on context. In the
> driver, though, the ultimate measure of correctness isn't a Nova call:
> it's the effect produced on the state of an external system.
>
> For the VMware driver we have nova.tests.virt.vmwareapi.fake. This is a
> lot of code: 1599 lines as of writing. It contains bugs, and it contains
> inaccuracies, and both of these can mess up tests. However:
>
> * It's vastly simpler than the system it models (vSphere server)
> * It's common code, so gets fixed over time
> * It allows tests to run almost all driver code unmodified
>
> So, for example, it knows that you can't move a file before you create
> it. It knows that creating a VM creates a bunch of different files, and
> where they're created. It knows what objects are created by the server,
> and what attributes they have. And what attributes they don't have. If
> you do an object lookup, it knows which objects to return, and what
> their properties are.
>
> All of this knowledge is vital to testing, and if it wasn't contained in
> the fake driver, or something like it[1], would have to be replicated
> across all tests which require it. i.e. It may be 1599 lines of
> complexity, but it's all complexity which has to live somewhere anyway.
>
> Incidentally, this is fresh in my mind because of
> https://review.openstack.org/#/c/122760/ . Note the diff stat: +70,
> -161, and the rewrite has better coverage, too :) It executes the
> function under test, it checks that it produces the correct outcome, and
> other than that it doesn't care how the function is implemented.
>
> TL;DR
>
> * Bootstrap hour is awesome
> * Don't fake if you don't have to
> * However, there are situations where it's a good choice

I'm going to push on this a bit further. Mocking is fine in many
cases, but it encodes dependencies from within your objects into the
test suite: e.g. that X will call Y and then Z - and if thats what you
want to *test*, thats great, but it often isn't outside of the very
smallest unit tests - because it makes your tests fragile to change -
tests for a class Foo need to change when a parent class Bar alters,
or when a utility class Quux alters. And that coupling has a cost and
is a pain.

Ad-hoc fakes are no better than mocks in this regard, but verified
fakes are substantially better and give nearly the same performance as
mocks, with (generally) better diagnostics, fidelity and low
fragility. Where we don't have verified fakes, making a single robust
fake for the interface is a good alternative - and I'd like to see
OpenStack do more of this, not less.

https://thoughtstreams.io/glyph/test-patterns/5952/ is a good post
from Glyph describing them.

-Rob

-- 
Robert Collins <rbtcollins at hp.com>
Distinguished Technologist
HP Converged Cloud