Proptest

Proptest is a property testing framework (i.e., the QuickCheck family) inspired by the Hypothesis framework for Python. It allows to test that certain properties of your code hold for arbitrary inputs, and if a failure is found, automatically finds the minimal test case to reproduce the problem. Unlike QuickCheck, generation and shrinking is defined on a per-value basis instead of per-type, which makes it more flexible and simplifies composition.

If you have dependencies which provide QuickCheck Arbitrary implementations, see also the related proptest-quickcheck-interop crates which enables reusing those implementations with proptest.

Status of this crate

The majority of the functionality offered by proptest is in active use and is known to work well.

The API is unlikely to see drastic breaking changes, but there may still be minor breaking changes here and there, particularly when "impl Trait" becomes stable and after the upcoming redesign of the rand crate.

See the changelog for a full list of substantial historical changes, breaking and otherwise.

Introduction

Property testing is a system of testing code by checking that certain properties of its output or behaviour are fulfilled for all inputs. These inputs are generated automatically, and, critically, when a failing input is found, the input is automatically reduced to a minimal test case.

Property testing is best used to compliment traditional unit testing (i.e., using specific inputs chosen by hand). Traditional tests can test specific known edge cases, simple inputs, and inputs that were known in the past to reveal bugs, whereas property tests will search for more complicated inputs that cause problems.

Getting Started

Let's say we want to make a function that parses dates of the form YYYY-MM-DD. We're not going to worry about validating the date, any triple of integers is fine. So let's bang something out real quick.

fn parse_date(s: &str) -> Option<(u32, u32, u32)> {
    if 10 != s.len() { return None; }
    if "-" != &s[4..5] || "-" != &s[7..8] { return None; }

    let year = &s[0..4];
    let month = &s[6..7];
    let day = &s[8..10];

    year.parse::<u32>().ok().and_then(
        |y| month.parse::<u32>().ok().and_then(
            |m| day.parse::<u32>().ok().map(
                |d| (y, m, d))))
}

It compiles, that means it works, right? Maybe not, let's add some tests.

#[test]
fn test_parse_date() {
    assert_eq!(None, parse_date("2017-06-1"));
    assert_eq!(None, parse_date("2017-06-170"));
    assert_eq!(None, parse_date("2017006-17"));
    assert_eq!(None, parse_date("2017-06017"));
    assert_eq!(Some((2017, 06, 17)), parse_date("2017-06-17"));
}

Tests pass, deploy to production! But now your application starts crashing, and people are upset that you moved Christmas to February. Maybe we need to be a bit more thorough.

In Cargo.toml, add

[dev-dependencies]
proptest = "0.8.7"

and at the top of main.rs or lib.rs:

#[macro_use] extern crate proptest;

Now we can add some property tests to our date parser. But how do we test the date parser for arbitrary inputs, without making another date parser in the test to validate it? We won't need to as long as we choose our inputs and properties correctly. But before correctness, there's actually an even simpler property to test: The function should not crash. Let's start there.

proptest! {
    #[test]
    fn doesnt_crash(s in "\\PC*") {
        parse_date(s);
    }
}

What this does is take a literally random &String (ignore \\PC* for the moment, we'll get back to that — if you've already figured it out, contain your excitement for a bit) and give it to parse_date() and then throw the output away.

When we run this, we get a bunch of scary-looking output, eventually ending with

thread 'main' panicked at 'Test failed: byte index 4 is not a char boundary; it is inside 'ௗ' (bytes 2..5) of `aAௗ0㌀0`; minimal failing input: s = "aAௗ0㌀0"
	successes: 102
	local rejects: 0
	global rejects: 0
'

If we look at the top directory after the test fails, we'll see a new proptest-regressions directory, which contains some files corresponding to source files containing failing test cases. These are failure persistence files. The first thing we should do is add these to source control.

$ git add proptest-regressions

The next thing we should do is copy the failing case to a traditional unit test since it has exposed a bug not similar to what we've tested in the past.

#[test]
fn test_unicode_gibberish() {
    assert_eq!(None, parse_date("aAௗ0㌀0"));
}

Now, let's see what happened... we forgot about UTF-8! You can't just blindly slice strings since you could split a character, in this case that Tamil diacritic placed atop other characters in the string.

In the interest of making the code changes as small as possible, we'll just check that the string is ASCII and reject anything that isn't.

fn parse_date(s: &str) -> Option<(u32, u32, u32)> {
    if 10 != s.len() { return None; }

    // NEW: Ignore non-ASCII strings so we don't need to deal with Unicode.
    if !s.is_ascii() { return None; }

    if "-" != &s[4..5] || "-" != &s[7..8] { return None; }

    let year = &s[0..4];
    let month = &s[6..7];
    let day = &s[8..10];

    year.parse::<u32>().ok().and_then(
        |y| month.parse::<u32>().ok().and_then(
            |m| day.parse::<u32>().ok().map(
                |d| (y, m, d))))
}

The tests pass now! But we know there are still more problems, so let's test more properties.

Another property we want from our code is that it parses every valid date. We can add another test to the proptest! section:

proptest! {
    // snip...

    #[test]
    fn parses_all_valid_dates(s in "[0-9]{4}-[0-9]{2}-[0-9]{2}") {
        parse_date(s).unwrap();
    }
}

The thing to the right-hand side of in is actually a regular expression, and s is chosen from strings which match it. So in our previous test, "\\PC*" was generating arbitrary strings composed of arbitrary non-control characters. Now, we generate things in the YYYY-MM-DD format.

The new test passes, so let's move on to something else.

The final property we want to check is that the dates are actually parsed correctly. Now, we can't do this by generating strings — we'd end up just reimplementing the date parser in the test! Instead, we start from the expected output, generate the string, and check that it gets parsed back.

proptest! {
    // snip...

    #[test]
    fn parses_date_back_to_original(y in 0u32..10000,
                                    m in 1u32..13, d in 1u32..32) {
        let (y2, m2, d2) = parse_date(
            &format!("{:04}-{:02}-{:02}", y, m, d)).unwrap();
        // prop_assert_eq! is basically the same as assert_eq!, but doesn't
        // cause a bunch of panic messages to be printed on intermediate
        // test failures. Which one to use is largely a matter of taste.
        prop_assert_eq!((y, m, d), (y2, m2, d2));
    }
}

Here, we see that besides regexes, we can use any expression which is a proptest::strategy::Strategy, in this case, integer ranges.

The test fails when we run it. Though there's not much output this time.

thread 'main' panicked at 'Test failed: assertion failed: `(left == right)` (left: `(0, 10, 1)`, right: `(0, 0, 1)`) at examples/dateparser_v2.rs:46; minimal failing input: y = 0, m = 10, d = 1
	successes: 2
	local rejects: 0
	global rejects: 0
', examples/dateparser_v2.rs:33
note: Run with `RUST_BACKTRACE=1` for a backtrace.

The failing input is (y, m, d) = (0, 10, 1), which is a rather specific output. Before thinking about why this breaks the code, let's look at what proptest did to arrive at this value. At the start of our test function, insert

    println!("y = {}, m = {}, d = {}", y, m, d);

Running the test again, we get something like this:

y = 2497, m = 8, d = 27
y = 9641, m = 8, d = 18
y = 7360, m = 12, d = 20
y = 3680, m = 12, d = 20
y = 1840, m = 12, d = 20
y = 920, m = 12, d = 20
y = 460, m = 12, d = 20
y = 230, m = 12, d = 20
y = 115, m = 12, d = 20
y = 57, m = 12, d = 20
y = 28, m = 12, d = 20
y = 14, m = 12, d = 20
y = 7, m = 12, d = 20
y = 3, m = 12, d = 20
y = 1, m = 12, d = 20
y = 0, m = 12, d = 20
y = 0, m = 6, d = 20
y = 0, m = 9, d = 20
y = 0, m = 11, d = 20
y = 0, m = 10, d = 20
y = 0, m = 10, d = 10
y = 0, m = 10, d = 5
y = 0, m = 10, d = 3
y = 0, m = 10, d = 2
y = 0, m = 10, d = 1

The test failure message said there were two successful cases; we see these at the very top, 2497-08-27 and 9641-08-18. The next case, 7360-12-20, failed. There's nothing immediately obviously special about this date. Fortunately, proptest reduced it to a much simpler case. First, it rapidly reduced the y input to 0 at the beginning, and similarly reduced the d input to the minimum allowable value of 1 at the end. Between those two, though, we see something different: it tried to shrink 12 to 6, but then ended up raising it back up to 10. This is because the 0000-06-20 and 0000-09-20 test cases passed.

In the end, we get the date 0000-10-01, which apparently gets parsed as 0000-00-01. Again, this failing case was added to the failure persistence file, and we should add this as its own unit test:

$ git add proptest-regressions

#[test]
fn test_october_first() {
    assert_eq!(Some(0, 10, 1), parse_date("0000-10-01"));
}

Now to figure out what's broken in the code. Even without the intermediate input, we can say with reasonable confidence that the year and day parts don't come into the picture since both were reduced to the minimum allowable input. The month input was not, but was reduced to 10. This means we can infer that there's something special about 10 that doesn't hold for 9. In this case, that "special something" is being two digits wide. In our code:

    let month = &s[6..7];

We were off by one, and need to use the range 5..7. After fixing this, the test passes.

The proptest! macro has some additional syntax, including for setting configuration for things like the number of test cases to generate. See its documentation for more details.

There is a more in-depth tutorial in the crate documentation.

Differences between QuickCheck and Proptest

QuickCheck and Proptest are similar in many ways: both generate random inputs for a function to check certain properties, and automatically shrink inputs to minimal failing cases.

The one big difference is that QuickCheck generates and shrinks values based on type alone, whereas Proptest uses explicit Strategy objects. The QuickCheck approach has a lot of disadvantages in comparison:

• QuickCheck can only define one generator and shrinker per type. If you need a custom generation strategy, you need to wrap it in a newtype and implement traits on that by hand. In Proptest, you can define arbitrarily many different strategies for the same type, and there are plenty built-in.

• For the same reason, QuickCheck has a single "size" configuration that tries to define the range of values generated. If you need an integer between 0 and 100 and another between 0 and 1000, you probably need to do another newtype. In Proptest, you can directly just express that you want a 0..100 integer and a 0..1000 integer.

• Types in QuickCheck are not easily composable. Defining Arbitrary and Shrink for a new struct which is simply produced by the composition of its fields requires implementing both by hand, including a bidirectional mapping between the struct and a tuple of its fields. In Proptest, you can make a tuple of the desired components and then prop_map it into the desired form. Shrinking happens automatically in terms of the input types.

• Because constraints on values cannot be expressed in QuickCheck, generation and shrinking may lead to a lot of input rejections. Strategies in Proptest are aware of simple constraints and do not generate or shrink to values that violate them.

The author of Hypothesis also has an article on this topic.

Of course, there's also some relative downsides that fall out of what Proptest does differently:

• Generating complex values in Proptest can be up to an order of magnitude slower than in QuickCheck. This is because QuickCheck performs stateless shrinking based on the output value, whereas Proptest must hold on to all the intermediate states and relationships in order for its richer shrinking model to work.

Limitations of Property Testing

Given infinite time, property testing will eventually explore the whole input space to a test. However, time is not infinite, so only a randomly sampled portion of the input space can be explored. This means that property testing is extremely unlikely to find single-value edge cases in a large space. For example, the following test will virtually always pass:

#[macro_use] extern crate proptest;
use proptest::prelude::*;

proptest! {
    #[test]
    fn i64_abs_is_never_negative(a: i64) {
        // This actually fails if a == i64::MIN, but randomly picking one
        // specific value out of 2⁶⁴ is overwhelmingly unlikely.
        assert!(a.abs() >= 0);
    }
}

Because of this, traditional unit testing with intelligently selected cases is still necessary for many kinds of problems.

Similarly, in some cases it can be hard or impossible to define a strategy which actually produces useful inputs. A strategy of .{1,4096} may be great to fuzz a C parser, but is highly unlikely to produce anything that makes it to a code generator.

Failure Persistence

By default, when Proptest finds a failing test case, it persists that failing case in a file named after the source containing the failing test, but in a separate directory tree rooted at proptest-regressions† . Later runs of tests will replay those test cases before generating novel cases. This ensures that the test will not fail on one run and then spuriously pass on the next, and also exposes similar tests to the same known-problematic input.

(† If you do not have an obvious source directory, you may instead find files next to the source files, with a different extension.)

It is recommended to check these files in to your source control so that other test runners (e.g., collaborators or a CI system) also replay these cases.

Note that, by default, all tests in the same crate will share that one persistence file. If you have a very large number of tests, it may be desirable to separate them into smaller groups so the number of extra test cases that get run is reduced. This can be done by adjusting the failure_persistence flag on Config.

There are two ways this persistence could theoretically be done.

The immediately obvious option is to persist a representation of the value itself, for example by using Serde. While this has some advantages, particularly being resistant to changes like tweaking the input strategy, it also has a lot of problems. Most importantly, there is no way to determine whether any given value is actually within the domain of the strategy that produces it. Thus, some (likely extremely fragile) mechanism to ensure that the strategy that produced the value exactly matches the one in use in a test case would be required.

The other option is to store the seed that was used to produce the failing test case. This approach requires no support from the strategy or the produced value. If the strategy in use differs from the one used to produce failing case that was persisted, the seed may or may not produce the problematic value, but nonetheless produces a valid value. Due to these advantages, this is the approach Proptest uses.

Forking and Timeouts

By default, proptest tests are run in-process and are allowed to run for however long it takes them. This is resource-efficient and produces the nicest test output, and for many use cases is sufficient. However, problems like overflowing the stack, aborting the process, or getting stuck in an infinite will simply break the entire test process and prevent proptest from determining a minimal reproducible case.

As of version 0.7.1, proptest has optional "fork" and "timeout" features (both enabled by default), which make it possible to run your test cases in a subprocess and limit how long they may run. This is generally slower, may make using a debugger more difficult, and makes test output harder to interpret, but allows proptest to find and minimise test cases for these situations as well.

To enable these features, simply set the fork and/or timeout fields on the Config. (Setting timeout implies fork.)

Here is a simple example of using both features:

#[macro_use] extern crate proptest;
use proptest::prelude::*;

// The worst possible way to calculate Fibonacci numbers
fn fib(n: u64) -> u64 {
    if n <= 1 {
        n
    } else {
        fib(n - 1) + fib(n - 2)
    }
}

proptest! {
    #![proptest_config(ProptestConfig {
        // Setting both fork and timeout is redundant since timeout implies
        // fork, but both are shown for clarity.
        fork: true,
        timeout: 1000,
        .. ProptestConfig::default()
    })]

    #[test]
    fn test_fib(n: u64) {
        // For large n, this will variously run for an extremely long time,
        // overflow the stack, or panic due to integer overflow.
        assert!(fib(n) >= n);
    }
}

The exact value of the test failure depends heavily on the performance of the host system, the rust version, and compiler flags, but on the system where it was originally tested, it found that the maximum value that fib() could handle was 39, despite having dozens of processes dump core due to stack overflow or time out along the way.

If you just want to run tests in subprocesses or with a timeout every now and then, you can do that by setting the PROPTEST_FORK or PROPTEST_TIMEOUT environment variables to alter the default configuration. For example, on Unix,

# Run all the proptest tests in subprocesses with no timeout.
# Individual tests can still opt out by setting `fork: false` in their
# own configuration.
PROPTEST_FORK=true cargo test
# Run all the proptest tests in subprocesses with a 1 second timeout.
# Tests can still opt out or use a different timeout by setting `timeout: 0`
# or another timeout in their own configuration.
PROPTEST_TIMEOUT=1000 cargo test

Acknowledgements

This crate wouldn't have come into existence had it not been for the Rust port of QuickCheck and the regex_generate crate which gave wonderful examples of what is possible.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.