Unordered thoughts about programming, engineering and dealing with the people in the process.

Neither Types Nor Tests Will Solve Your Data Coverage Problem

| Comments

I just watched the Strangeloop talk titled Types vs. Tests: An Epic Battle from Amanda Laucher and Paul Snively. As Amanda says it’s a discussion many of us have had in the past, I used to talk about it with fedesilva, hardcore Scala advocate, you know what I mean?

For me, I think types and tests are two sides of the same coin, because neither strategy for proving correctness is computable.

Bear with me.

The purpose of having types or having tests, it’s to prove your program to be correct before your bugs reach your customers, one strategy tries to prove it at compile time and the other after compile time. But what it really means your program to be correct?

Let’s go for ride on Computability Theory.

Functions and programs

According to the definition a function f is computable if a program P exists which can calculate the function given unlimited amounts of time and storage.

f:N→N is computable ↔ ∃ a program P which computes the function.

Also we must define when a program P converges for input n.

Program P with input n converges if ∃ m ∈ N / <Q,n> = m , it's written <Q,n>↓

In computer theory there are well known functions to be non computable, two of them are:

Θ(n) = 1  if <Ix(n), n>↓
       0  if <Ix(n), n>↑

Which says that the function Θ is equal to 1 if the program of index n converges on input n and 0 if the program of index n diverges on input n.

There’s another very famous function which has been proved to be non computable.

stop(p, n) = 1 if <Ix(p), n>↓
             0 if <Ix(p), n>↑

Which pretty much says that given a description of an arbitrary computer program, decide whether the program finishes running or continues to run forever. As you may have guessed, it’s the well known Halting Problem, and it’s not computable.

How the Halting Problem relates to types and tests

Let’s assume our program P we’re trying to prove correct, computes the function f:N→N. We can define our program that proves correctness, a program T that computes the function


Which is to say, for every input of the domain, our program T decides if the program converges or not. It’s starting to sound familiar ain’t it?

Let’s assume such a program T exists to prove correctness, and we have a macro MT to find such a program given P. We could write the following program Q

PROGRAM(x0, x1)
x4=EVAL_PROG(x3, x1)

What Q does it’s having received a program x0 and an input x1, first finds the T decider program for the program x0, and then evaluates the program with the input x1.

So what do we have here?

Q(p, n) = 1 ↔ ft(n) = 1 ↔ <P, n> ↓

So we have written a program which computes the stop function, which is absurd. It means we cannot have a program that decides on the computability of a program.

Show me the code

In practice, it means that if you have this program

int f(int x) {
   while (x > 0) {
   return 0;

This program doesn’t stop for x < 0, and according to theory, there’s no program you can write to find out about it.

There are also a few other funny cases regarding the domain of your functions, such as

int f(int x) {
   return 1 / (x - 3);

This function fails miserably if x = 3. Just think about it when your functions have a more complex domain.

How to improve your tests for correctness

Most people I see are worried about having 100% code coverage, but it’s not that usual to see people worried about data coverage.

As seen in the previous example if you forget to test for x = 3 you may have 100% code coverage but your program will blow up anyway.

Regarding types, I know Dependent Types exists, but it’s the other side of the same coin, you have to provide a constructive proof that the type is inhabited. So if you don’t define your type considering the special cases of your function domain, no one is coming up to save your ass.

But when thinking about correctness you should be thinking about your function domain.


Both tests and types are useful ways to validate your program is correct, but not perfect. Even the discussion is meaningless, because it’s just a matter of taste whether you like to specify your correctness rules in types or tests, but it’s something you will keep doing as far as I can tell.

As Rich Hickey said, both tests and types are like guard rails, and you must know the cliff is there in order to decide building them.

Update: Many people wrote to me as if I’m saying you can’t prove a program to be correct, that was not what I’ve tried to say. It was that you can’t have a system that can prove programs to be correct without specifying the rules yourself. That is, it was a case against Q not against T.

And hey! you follow me on twitter!