Hypothesis

In my last post I mentioned the problem of bug slippage: When you start with one bug, reduce the test case, and end up with another bug.

I’ve run into another related problem twice now, and it’s not one I’ve seen talked about previously.

The problem is this: Sometimes shrinking makes a bug seem much less interesting than it actually is.

When Hypothesis finds an example triggering a bug, it tries to shrink the example down to something simpler that triggers it. This is a pretty common feature, and most property-based testing libraries implement something similar (though there are a number of differences between them). Stand-alone test case reducers are also fairly common, as it’s a useful thing to be able to do when reporting bugs in external projects - rather than submitting a giant file triggering the bug, a good test case reducer can often shrink it down to a couple of lines.

But there’s a problem with doing this: How do you know that the bug you started with is the same as the bug you ended up with?

This isn’t just an academic question. It’s very common for the bug you started with to slip to another one.

Consider for example, the following test:

from hypothesis import given, strategies as st

def mean(ls):
    return sum(ls) / len(ls)


@given(st.lists(st.floats()))
def test(ls):
    assert min(ls) <= mean(ls) <= max(ls)

This has a number of interesting ways to fail: We could pass NaN, we could pass [-float('inf'), +float('inf')], we could pass numbers which trigger a precision error, etc.

But after test case reduction, we’ll pass the empty list and it will fail because we tried to take the min of an empty sequence.

This isn’t necessarily a huge problem - we’re still finding a bug after all (though in this case as much in the test as in the code under test) - and sometimes it’s even desirable - you find more bugs this way, and sometimes they’re ones that Hypothesis would have missed - but often it’s not, and an interesting and rare bug slips to a boring and common one.

Historically Hypothesis has had a better answer to this than most - because of the Hypothesis example database, all intermediate bugs are saved and a selection of them will be replayed when you rerun the test. So if you fix one bug then rerun the test, you’ll find the other bugs that were previously being hidden from you by that simpler bug.

But that’s still not a great user experience - it means that you’re not getting nearly as much information as you could be, and you’re fixing bugs in Hypothesis’s priority order rather than yours. Wouldn’t it be better if Hypothesis just told you about all of the bugs it found and you could prioritise them yourself?

Well, as of Hypothesis 3.29.0, released a few weeks ago, now it does!

If you run the above test now, you’ll get the following:

Falsifying example: test(ls=[nan])
Traceback (most recent call last):
  File "/home/david/hypothesis-python/src/hypothesis/core.py", line 671, in run
    print_example=True, is_final=True
  File "/home/david/hypothesis-python/src/hypothesis/executors.py", line 58, in default_new_style_executor
    return function(data)
  File "/home/david/hypothesis-python/src/hypothesis/core.py", line 120, in run
    return test(*args, **kwargs)
  File "broken.py", line 8, in test
    def test(ls):
  File "/home/david/hypothesis-python/src/hypothesis/core.py", line 531, in timed_test
    result = test(*args, **kwargs)
  File "broken.py", line 9, in test
    assert min(ls) <= mean(ls) <= max(ls)
AssertionError

Falsifying example: test(ls=[])
Traceback (most recent call last):
  File "/home/david/hypothesis-python/src/hypothesis/core.py", line 671, in run
    print_example=True, is_final=True
  File "/home/david/hypothesis-python/src/hypothesis/executors.py", line 58, in default_new_style_executor
    return function(data)
  File "/home/david/hypothesis-python/src/hypothesis/core.py", line 120, in run
    return test(*args, **kwargs)
  File "broken.py", line 8, in test
    def test(ls):
  File "/home/david/hypothesis-python/src/hypothesis/core.py", line 531, in timed_test
    result = test(*args, **kwargs)
  File "broken.py", line 9, in test
    assert min(ls) <= mean(ls) <= max(ls)
ValueError: min() arg is an empty sequence

You can add @seed(67388524433957857561882369659879357765) to this test to reproduce this failure.
Traceback (most recent call last):
  File "broken.py", line 12, in <module>
    test()
  File "broken.py", line 8, in test
    def test(ls):
  File "/home/david/hypothesis-python/src/hypothesis/core.py", line 815, in wrapped_test
    state.run()
  File "/home/david/hypothesis-python/src/hypothesis/core.py", line 732, in run
    len(self.falsifying_examples,)))
hypothesis.errors.MultipleFailures: Hypothesis found 2 distinct failures.

(The stack traces are a bit noisy, I know. We have an issue open about cleaning them up).

All of the different bugs are minimized simultaneously and take full advantage of Hypothesis’s example shrinking, so each bug is as easy (or hard) to read as if it were the only bug we’d found.

This isn’t perfect: The heuristic we use for determining if two bugs are the same is whether they have the same exception type and the exception is thrown from the same line. This will necessarily conflate some bugs that are actually different - for example, [float('nan')], [-float('inf'), float('inf')] and [3002399751580415.0, 3002399751580415.0, 3002399751580415.0] each trigger the assertion in the test, but they are arguably “different” bugs.

But that’s OK. The heuristic is deliberately conservative - the point is not that it can distinguish whether any two examples are the same bug, just that any two examples it distinguishes are different enough that it’s interesting to show both, and this heuristic definitely manages that.

As far as I know this is a first in property-based testing libraries (though something like it is common in fuzzing tools, and theft is hot on our tail with something similar) and there’s been some interesting related but mostly orthogonal research in Erlang QuickCheck.

It was also surprisingly easy.

A lot of things went right in writing this feature, some of them technical, some of them social, somewhere in between.

The technical ones are fairly straightforward: Hypothesis’s core model turned out to be very well suited to this feature. Because Hypothesis has a single unified intermediate representation which defines a total ordering for simplicity, adapting Hypothesis to shrink multiple things at once was quite easy - whenever we attempt a shrink and it produces a different bug than the one we were looking for, we compare it to our existing best example for that bug and replace it if the current one is better (or we’ve discovered a new bug). We then just repeatedly run the shrinking process for each bug we know about until they’ve all been fully shrunk.

This is in a sense not surprising - I’ve been thinking about the problem of multiple-shrinking for a long time and, while this is the first time it’s actually appeared in Hypothesis, the current choice of model was very much informed by it.

The social ones are perhaps more interesting. Certainly I’m very pleased with how they turned out here.

The first is that this work emerged tangentially from the recent Stripe funded work - Stripe paid me to develop some initial support for testing Pandas code with Hypothesis, and I observed a bunch of bug slippage happening in the wild while I was testing that (it turns out there are quite a lot of ways to trigger exceptions from Pandas - they weren’t really Pandas bugs so much as bugs in the Pandas integration, but they still slipped between several different exception types), so that was what got me thinking about this problem again.

Not by accident, this feature also greatly simplified the implementation of the new deadline feature that Smarkets funded, which was going to have to have a lot of logic about how deadlines and bugs interacted, but all that went away as soon as we were able to handle multiple bugs sensibly.

This has been a relatively consistent theme in Hypothesis development - practical problems tend to spark related interesting theoretical developments. It’s not a huge exaggeration to say that the fundamental Hypothesis model exists because I wanted to support testing Django nicely. So the recent funded development from Stripe and Smarkets has been a great way to spark a lot of seemingly unrelated development and improve Hypothesis for everyone, even outside the scope of the funded work.

Another thing that really helped here is our review process, and the review from Zac in particular.

This wasn’t the feature I originally set out to develop. It started out life as a much simpler feature that used much of the same machinery, and just had a goal of avoiding slipping to new errors all together. Zac pushed back with some good questions around whether this was really the correct thing to do, and after some experimentation and feedback I eventually hit on the design that lead to displaying all of the errors.

Our review handbook emphasises that code review is a collaborative design process, and I feel this was a particularly good example of that. We’ve created a great culture of code review, and we’re reaping the benefits (and if you want to get in on it, we could always use more people able and willing to do review…).

All told, I’m really pleased with how this turned out. I think it’s a nice example of getting a lot of things right up front and this resulting in a really cool new feature.

I’m looking forward to seeing how it behaves in the wild. If you notice any particularly fun examples, do let me know, or write up a post about them yourself!

If you look at the original property-based testing library, the Haskell version of QuickCheck, tests are very closely tied to types: The way you typically specify a property is by inferring the data that needs to be generated from the types the test function expects for its arguments.

This is a bad idea.

I’m in the process of trying to turn my work on Hypothesis into a PhD and I realised that I don’t have a good self-contained summary as to why researchers should care about it.

So this is that piece. I’ll try to give a from scratch introduction to the why and what of Hypothesis. It’s primarily intended for potential PhD supervisors, but should be of general interest as well (especially if you work in this field).

Why should I care about Hypothesis from a research point of view?

The short version:

Hypothesis takes an existing effective style of testing (property-based testing) which has proven highly effective in practice and makes it accessible to a much larger audience. It does so by taking several previously unconnected ideas from the existing research literature on testing and verification, and combining them to produce a novel implementation that has proven very effective in practice.

The long version is the rest of this article.

Hypothesis has a very different underlying implementation to any other property-based testing system. As far as I know, it’s an entirely novel design that I invented.

Central to this design is the following feature set which every Hypothesis strategy supports automatically (the only way to break this is by having the data generated depend somehow on external global state):

1. All generated examples can be safely mutated
2. All generated examples can be saved to disk (this is important because Hypothesis remembers and replays previous failures).
3. All generated examples can be shrunk
4. All invariants that hold in generation must hold during shrinking ( though the probability distribution can of course change, so things which are only supported with high probability may not be).

(Essentially no other property based systems manage one of these claims, let alone all)

The initial mechanisms for supporting this were fairly complicated, but after passing through a number of iterations I hit on a very powerful underlying design that unifies all of these features.

It’s still fairly complicated in implementation, but most of that is optimisations and things needed to make the core idea work. More importantly, the complexity is quite contained: A fairly small kernel handles all of the complexity, and there is little to no additional complexity (at least, compared to how it normally looks) in defining new strategies, etc.

This article will give a high level overview of that model and how it works.

In my last article about shrinking, I discussed the problems with basing shrinking on the type of the values to be shrunk.

In writing it though I forgot that there was a halfway house which is also somewhat bad (but significantly less so) that you see in a couple of implementations.

This is when the shrinking is not type based, but still follows the classic shrinking API that takes a value and returns a lazy list of shrinks of that value. Examples of libraries that do this are theft and QuickTheories.

This works reasonably well and solves the major problems with type directed shrinking, but it’s still somewhat fragile and importantly does not compose nearly as well as the approaches that Hypothesis or test.check take.

Ideally, as well as not being based on the types of the values being generated, shrinking should not be based on the actual values generated at all.

This may seem counter-intuitive, but it actually works pretty well.

One of the big differences between Hypothesis and Haskell QuickCheck is how shrinking is handled.

Specifically, the way shrinking is handled in Haskell QuickCheck is bad and the way it works in Hypothesis (and also in test.check and EQC) is good. If you’re implementing a property based testing system, you should use the good way. If you’re using a property based testing system and it doesn’t use the good way, you need to know about this failure mode.

Unfortunately many (and possibly most) implementations of property based testing are based on Haskell’s QuickCheck and so make the same mistake.

This is one of the most common first questions about Hypothesis.

People generally assume that the number of tests run will depend on the specific strategies used, but that’s generally not the case. Instead Hypothesis has a fairly fixed set of heuristics to determine how many times to run, which are mostly independent of the data being generated.

But how many runs is that?

The short answer is 200. Assuming you have a default configuration and everything is running smoothly, Hypothesis will run your test 200 times.

The longer answer is “It’s complicated”. It will depend on the exact behaviour of your tests and the value of some settings. In this article I’ll try to clear up some of the specifics of which settings affect the answer and how.

What happens when you run a test using Hypothesis? This article will help you understand.