This package is deprecated. Use
elm-test
's fuzz tests instead.
Traditional unit-testing consists in asserting that certain inputs yield certain outputs. Property-based testing makes
claims relating input and output. These claims can then be automatically tested over as many randomly-generated inputs
as desired. If a failing input is found, it can be "shrunk" to compute a minimal failing case which is more
representative of the bug. The goal of elm-check
is to automate this process.
This library can replace many unit tests, but it cannot test asynchronous, UI, or end-to-end functionality.
Suppose you wanted to test List.reverse
. A correct implementation will obey a number of properties (or assertions),
regardless of the list being reversed, including:
- Reversing a list twice yields the original list.
- Reversing does not modify the length of a list.
You can make these claims in elm-check
as follows:
myClaims : Claim
myClaims =
suite "List Reverse"
[ claim
"Reversing a list twice yields the original list"
`that`
(\list -> reverse (reverse list))
`is`
identity
`for`
list int
, claim
"Reversing a list does not modify its length"
`that`
(\list -> length (reverse list))
`is`
(\list -> length list)
`for`
list int
]
As, you can see, elm-check
defines a Domain-Specific Language (DSL) for writing claims. It may look odd at first, but
the code is actually very straightforward to work with.
Straightforward?! It might help to review some language features being used. First,
suite
takes a string and a list, which forms most of the code. The list actually has only two items, the result of callingclaim
twice. (See the comma right before the secondclaim
?) Backticks indicate that a function is being called infix.(\x -> thing x)
is an anonymous function.
Let's examine each component of a claim.
claim <String>
This is the name of the test and is used when output is displayed, so make it descriptive.that <function>
This is the "actual" value, the result of the code or feature under test.is <function>
This is the "expected" value. Think of it like a control in a science experiment. It's the value that isn't complicated. A test claims that, for any inputx
,actual x == expected x
.for <Producer>
AnProducer
is basically a way to randomly create values for the inputs to the functions. So rather than operating on a single example, like unit testing, it can test that a relationship holds for many values. There's an entire module full ofProducer
s so you can test almost anything.
We also group our two claims into a suite. Suites can be nested within other suites as deep as you like, so they're useful for organizing tests of many features or modules.
Once you've built your claims, verifying them is easy:
evidence : Evidence
evidence = quickCheck myClaims
quickCheck
will take either a single claim or a suite of claims and will run 100 checks on each claim to attempt to
disprove each claim. quickCheck
will then return a descriptive result of the checks performed, in the Evidence
type.
You can dive into these results if you like, but the simplest way to know "did my tests pass" is to use elm-test.
main = ElmTest.elementRunner (Check.Test.evidenceToTest evidence)
Running the page in elm reactor
will inform you that all tests have passed. (You can find the complete code under
examples
.)
Suppose you start with a number x
. Mathematically, if you multiply by another number y
, and then divide by y
, you
should be left with x
. You would make this claim as follows:
myClaims =
claim
"Multiplication and division are inverse operations"
`that`
(\(x, y) -> x * y / y)
`is`
(\(x, y) -> x)
`for`
tuple (float, float)
Note that we're using the tuple
producer because the functions we pass must take exactly one argument. If you put
this into the program above, you'd get:
Multiplication and division are inverse operations: FAILED.
On check 23, found counterexample: (0,0)
Expected: 0
But It Was: NaN
This result shows that elm-check
has found a counter example, namely (0,0)
which falsifies the claim. This is
obviously true because division by 0 is undefined, hence the NaN
value.
We can easily exclude zero by filtering the producer. Change the last line to:
filter (\(x, y) -> y /= 0) (tuple (float, float))
This function (in Check.Producer
) will only use values that meet our criteria (not being equal to zero). This is
preferable to changing the expected and actual functions because it's simpler, and it doesn't reduce the number of
inputs we try.
Now we get a different error.
Multiplication and division are inverse operations, if zero is omitted: FAILED.
On check 20, found counterexample: (0.00019869294196802492,0.0001670854544888915)
Expected: 0.00019869294196802492
But It Was: 0.00019869294196802494
Floating point arithmetic strikes again! Notice that the expect and the actual values only differ by a tiny amount.
Instead of claiming equality, we want to claim that the two values are near to each other. In particular, we want to say that the difference of these values is very close to zero. Rather than supplying expected and actual, we will supply a function that we expect to always be true.
myClaims : Claim
myClaims =
claim
"Multiplication and division are near inverse operations"
`true`
(\(x, y) -> abs ((x * y / y) - x) < 1e-6)
`for`
filter (\(x, y) -> y /= 0) (tuple (float, float))
The test now passes. This gives us confidence that multiplication and division are very nearly inverses, for any pair of floats where the second one isn't zero.
The DSL can give difficult error messages. Ensure that each claim uses one of these three patterns:
- claim - (string) - that - (actual) - is - (expected) - for - (producer)
- claim - (string) - true - (predicate) - for - (producer)
- claim - (string) - false - (predicate) - for - (producer)
Ensure that each of these words except claim
is surrounded by backticks.
If you're putting main claims together in a suite, ensure that you have commas between each claim.
Ensure that the two functions you pass have the same type. Ensure the input type matches the producer. Ensure the output type is something equatable -- functions aren't, so be sure you fully apply them.
It can be difficult to write claims about a system, especially if it's not simple mathematics or a data structure.
Jessica Kerr suggests writing "a box around the API". Rather than specifying an expected value exactly, you should try to indicate a range in which it can reasonably fall.
You may have noticed in the division example that the second pair of failing values were both very close to zero. This is because of a process called shrinking, which in the case of floats, happens to bring them closer to zero. It makes lists, strings, and most other things smaller.
Here's how it works, when elm-check
encounters a failing test, it has strategies to shrink the input that caused the
failure. If any of those inputs cause a failure, it tries to shrink them in turn, until it has found a minimal failing
test case. Small examples of failure tend to be much more helpful for debugging.
Here's the thing: all of this happens automatically. You get smaller, easier-to-understand counterexamples, for free.
We used the quickCheck
function above to run our tests. There is also check
, which allows you to supply a random seed
and specify the number of tests to run per claim, in case you think 100 is insufficient. More tests increase the
likelihood of finding obscure bugs, but take longer.
Once again, the easiest way to view the results of your tests is Check.Test.evidenceToTest
. The resulting value can be
used with any of elm-test
's runners, including on the console for CI builds.
If you really want to explore the results of your tests, the Evidence
type is fully exposed and includes a large
amount of information.
You may want to test a function whose input does not have an producer available. If possible, convert or map over an
existing producer to obtain the one you need. If necessary, you can write your own because the definition of Producer
is exported. You'll need to dive into elm-shrink
, as well and the Random
module.
The Investigator
type has been renamed Producer
. You should do a find-and-replace. If you defined your own
investigators, you'll need to use the type alias directly. So investigator generator shrinker
should become Producer generator shrinker
. keepIf
and dropIf
have been changed to filter
. void
is now unit
.
The arguments to check
have been reordered so that the Claim
is last.
If you relied on claimN
, claimNTrue
, and so on, you will need to rewrite your tests in the DSL. If you used the DSL
in Check.Test
, you will need to rewrite your tests using the main DSL, and then use Check.Test.evidenceToTest
for
Integration with elm-test
.