Finding Bugs with Property Based Testing in a Statistics Calculation

I always loved the idea of Property Based Testing, it was the most mind blowing idea I encountered while learning clojure many years back. However, I always found it hard to apply in practice. I barely encountered the “easy” case where an operation was reversible (if I encode a term and then decode it again it should be equal to the original term, for instance). And the properties I came up always seemed to loose to catch any bugs.

So, I thought well let’s give it another shot. I got all these statistics functions in benchee/statistex – data should be easy to generate for them. Let’s give it a shot, we’ll surely not find a bug… or will we?

The first property based tests

If nothing else, I wanted to make sure no matter what numbers you throw at my statistics module it won’t blow up. To implement it I used stream_data:

This is what I came up with – our samples are any non empty list of floats and there are a bunch of checks that make sure the values are somewhere between minimum and maximum or bigger than 0. No way the tests are failing…

Wait, the tests are failing?!

 Failed with generated values (after 2 successful runs):

* Clause: samples <- list_of(float(), min_length: 1) Generated: [-9.0, -1.0] Assertion with >= failed
code: assert stats.standard_deviation_ratio() >= 0
left: -1.131370849898476
right: 0

Honestly, I was shocked. On closer inspection, the standard deviation ratio was negative when I said it should always be positive. As the generated sample only contains negative numbers the average is negative as well. As the ratio is calculated by dividing the standard deviation by the average it turned out to be negative. Usually I only work with positive samples, hence it never occurred before. The ratio should still always be positive so an abs/1 call fixed it.

Another property

Thinking more I came up with another property:

It’s much like the first properties, just making sure the percentile values are in order as they should there is absolutely no possibility that this will fail, absolutely none, well tested code… no chance it will fail …

IT FAILED AGAIN?!?!?!

 Failed with generated values (after 4 successful runs):

* Clause: samples <- list_of(float(), min_length: 1)
Generated: [1.0, 33.0]

Assertion with <= failed
code: assert percies[25] <= percies[50]
left: 25.0
right: 17.0

Wait, the 25th percentile is bigger than the 50th percentile? No way that’s ok.

A lot of digging, googling and reading our original source for implementing percentile interpolation later I figured out the problem. Basically interpolation for small sample sizes is hard and also uncommon. We missed a clause/case stated in the source, that points out that for a too small percentile and sample size the value is to be set to the minimum.

Note that any p ≤ 1/(N+1) will simply be set to the minimum value.

Our p was 0.25 (25th percentile) and 0.25 <= 1/3. Implementing this clause (through guard clauses) fixed the test.

You can check out the full implementation and bug fixes in the PR.

Learnings

The generation part was super easy in the case shown. However, what’s impressive to me is that although the properties were very loosely defined they still uncovered 2 bugs. And that’s in code that both me and many of you have been running for quite a while in benchee. Sure, they are very specific edge cases but that’s what property based testing is good at: Finding edge cases!

If you have other ideas for properties to check, I’m happy to listen and learn. And give property based testing a shot yourselves even with very loose properties – you might be surprised what you find.

 

One thought on “Finding Bugs with Property Based Testing in a Statistics Calculation

  1. Thanks for this posting. It is a good and needed lesson to all of us supposedly professional software developers who do not avail ourselves of the wide range of available testing technologies. It is too easy to spend a few minutes writing the obvious tests and then sit back and think “yep that should cover it…I need to move on”. Um, think again!

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s