Cosmic Ray: mutation testing for Python

"Four human beings -- changed by space-born cosmic rays into something more than merely human." — The Fantastic Four

Cosmic Ray is a tool for performing mutation testing on Python code.

N.B.! Cosmic Ray is still learning how to walk!

At this time Cosmic Ray is young and incomplete. It doesn't support all of the mutations it should, its output format is crude, it only supports some forms of test discovery, it may fall over on exotic modules...the list goes on and on. Still, for the adventurous it does work. Hopefully things will improve fairly rapidly.

And, of course, patches and ideas are welcome.


If you just want to get down to the business of finding and killing mutants, here's what you do:

  1. Install Cosmic Ray

    pip install cosmic_ray

  2. Initialize a Cosmic Ray session

    cosmic-ray init --baseline=10 <session name> <module name> -- <test directory>

  3. Execute the session:

    cosmic-ray exec <session name>

  4. View the results:

    cosmic-ray report <session name>

This will print out a bunch of information about what Cosmic Ray did, including what kinds of mutants were created, which were killed, and – chillingly – which survived.

A concrete example: running the adam unittests

Cosmic Ray includes a number of unit tests which perform mutations against a simple module called adam. As a way of test driving Cosmic Ray, you can run these tests, too, like this:

cd test_project
cosmic-ray init --baseline=10 example-session adam -- tests
cosmic-ray --verbose exec example-session
cosmic-ray report example-session

In this case we're passing the --verbose flag to the exec command so that you can see what Cosmic Ray is doing. If everything goes as expected, the report command will report a 0% survival rate.


You can install Cosmic Ray using pip:

pip install cosmic_ray

Or you can use the supplied

python install

Both of these approaches will install the Cosmic Ray package and create an executable called cosmic-ray.

Virtual environments

You'll often want to install Cosmic Ray into a virtual environment. However, you generally don't want to install it into its own. Rather, you want to install it into the virtual environment of the project you want to test. This ensures that the test runners have access to the modules they are supposed to test.


Cosmic Ray has a notion of sessions which encompass an entire mutation testing run. Essentially, a session is a database which records the work that needs to be done for a run. Then as results are available from workers that do the actual testing, the database is updated with results. By having a database like this, Cosmic Ray can safely stop in the middle of a (potentially very long) session and be restarted. Since the session knows which work is already completed, it can continue where it left off.

Sessions also allow for arbitrary post-facto analysis and report generation.

Initializing sessions

Before you can do mutation testing with Cosmic Ray, you need to first initialize a session. You can do this using the init command. With this command you tell Cosmic Ray a) the name of the session, b) which module(s) you wish to mutate and c) the location of the test suite. For example, if you've a package named allele and if the unittest tests for the package are all under the directory allele_tests, you would run cosmic-ray init like this:

cosmic-ray init --baseline=2 test_session allele -- allele_tests

You'll notice that this creates a new file called "test_session.json". This the database for your session.

There are a number of other options you can pass to the init command; see the help message for more details.

An important note on separating tests and production code

Cosmic Ray has a relatively simple view of how to mutate modules. Fundamentally, it will attempt to mutate any and all code in a module. This means that if you have test code in the same module as your code under test, Cosmic Ray will happily mutate the test code along with the production code. This is probably not what you want.

The best way to avoid this problem is to keep your test code in separate modules from your production code. This way you can tell Cosmic Ray precisely what to mutate.

Ideally, your test code will be in a different package from your production code. This way you can tell Cosmic Ray to mutate an entire package without needing to filter anything out. However, if your test code is in the same package as your production code (a common configuration), you can use the --exclude-modules flag of cosmic-ray init to prevent mutation of your tests.

Given the choice, though, we recommend keeping your tests outside of the package for your code under test.

Executing tests

Once a session has been initialized, you can start executing tests by using the exec command. This command just needs the name of the session you provided to init:

cosmic-ray exec test_session

Normally this won't produce any output unless there are errors.

Viewing the results

Once your tests have completed, you can view the results using the report command:

cosmic-ray report test-session

This will give you detailed information about what work was done, followed by a summary of the entire session.

Short-cut: the run command

Originally Cosmic Ray didn't have a notion of sessions, and didn't distinguish between initialization and execution of the tests. It did all of its work using the run command.

Recent versions of Cosmic Ray still support the run command. All this command does is first do an init followed by an exec. This can be convenient for small test runs.

Be aware, however, that init can destroy an existing session database! If you've got a session database with results representing hours of execution, you probably don't want to delete it! So be aware that using the init or run command have the potential to delete data.

Test runners

Cosmic Ray supports multiple test runners. A test runner is simply a plugin that supports a particular way of running tests. For example, there is a test runner for tests written with the standard unittest module, and there's another for tests written using py.test.

To specify a particular test runner when running Cosmic Ray, pass the --test-runner flag to the init subcommand. For example, to use the pytest runner you would use:

cosmic-ray init --test-runner=pytest test_session allele -- allele_tests

To get a list of the available test runners, use the test-runners subcommand:

cosmic-ray test-runners

Test runners require information about which tests to run, flags controlling their behavior, and so forth. Since each test runner implementation takes different kinds of information, we allow users to pass arbitrary lists of arguments to test runners. When running the cosmic-ray init command, everything after the lone -- token is passed verbatim to the test runner initializer.

For example, the command:

cosmic-ray init --test-runner=pytest sess allele -- -x -k test_foo allele_tests

would pass the list ['-x', '-k', 'test_foo', 'allele_tests'] to the pytest runner initializer. This plugin passes this list directly to the pytest.main() function which treats them as command line arguments; in this case, it means "exit on first failure, only running tests under 'allele_tests' which match 'test_foo'". Each test runner will accept different arguments, so see their documentation for details on how to use them.

Specifying test timeouts

One difficulty mutation testing tools have to face is how to deal with mutations that result in infinite loops (or other pathological runtime effects). Cosmic Ray takes the simple approach of using a timeout to determine when to kill a test and consider it incompetent. That is, if a test of a mutant takes longer than the timeout, the test is killed, and the mutant is marked incompetent.

There are two ways to specify timeout values to Cosmic Ray. The first is through the --timeout flag for the init subcommand. This flags specifies an absolute number of seconds that a test will be allowed to run. After the timeout is up, the test is killed. For example, to specify that tests should timeout after 10 seconds, use:

cosmic-ray init --timeout=10 test_session allele -- allele/tests

The second way is by using a baseline timing. To use this technique, pass the --baseline argument to the init subcommand. When Cosmic Ray sees this flag it will make an initial run of the tests on an un-mutated version of the module under test. The amount of time this takes is considered the baseline timing. Then, Cosmic Ray multiplies this baseline timing by the value of --baseline and this final value is used as the timeout for tests. For example, to tell Cosmic Ray to timeout tests when they take 3 times longer than a baseline run, use:

cosmic-ray init --baseline=3 test_session allele -- allele/tests

This baseline technique is particularly useful if your testsuite runtime is in flux.

Running with a config file

For many projects you'll probably be running the same cosmic-ray command over and over. Instead of having to remember and retype potentially complex commands each time, you can store cosmic-ray commands in a config file. You can then execute these commands by passing the load command to cosmic-ray.

Each line in the config file is treated as a separate command-line argument to cosmic-ray. Empty lines in the file are skipped, and you can have comments in config files that start with #.

So, for example, if you need to invoke this command for your project:

cosmic-ray run --verbose --timeout=30 --no-local-import --baseline=2 allele -- allele/tests/unittests

you could instead create a config file, cr-allele.conf, with these contents:

--verbose     # this can be useful for debugging
--timeout=30  # this is plenty of time

Then to run the command in that config file:

cosmic-ray load cr-allele.conf

and it will have the same effect as running the original command.

Distributed testing with Celery

One of the main practical challenges to mutation testing is that it can take a long time. Even on moderately sized projects, you might need millions of individual mutations and test runs. This can be prohibitive to run on a single system.

One way to cope with these long runtimes is to parallelize the mutation and testing procedures. Fortunately, mutation testing is embarassingly parallel in nature, so we can apply some relatively simple techniques to get really nice scaling up of the work. We've chosen to use the Celery distributed task queue to spread work across multiple nodes.

The basic idea is very simple. Celery lets you start multiple workers which listen for commands from a task queue. A central process creates all of the commands for a mutation testing run, and these commands are distributed to the workers as they become available. When a worker receives a command, it starts a new python process (using the worker subcommand to Cosmic Ray) which performs a single mutation and runs the test suite.

Spawning a separate process for each test suite may seem expensive. However, it's the best way we have for ensuring that pathological mutants can't somehow corrupt the runtime of the worker processes. And ultimately the cost of starting the process is likely to be very small compared to the runtime of the test suite.

By its nature, Celery lets you start workers on as many systems as you want, all connected to the same task queue. So you could potentially have thousands of workers performing mutation testing runs, giving nearly perfect scaling! While not everyone has thousands of machines on hand to do their testing work, it's conceivable that Cosmic Ray will one day be able to work with machines on commodity cloud providers, meaning that highly-scaled mutation testing for Python will be available to anyone who wants it.

Installing RabbitMQ

Celery is primarily a Python API atop the RabbitMQ task queue. As such, if you want to use Cosmic Ray in distributed mode you first need to install RabbitMQ and run the server. The steps for installing and running RabbitMQ are covered in detail at that project's site, so go there for more information. Make sure the RabbitMQ server is installated and running before going any further with distributed execution.

Starting distributed worker processes

Once RabbitMQ is running, you need to start some worker processes which will do the actualy mutation testing. Start one or more worker processes like this:

celery -A cosmic_ray.tasks.worker worker

You should do this, of course, from the virtual environment into which you've installed Cosmic Ray. Similary, you need to make sure that the worker is in an environment in which it can import the modules under test. Generally speaking, you can meet both of these criteria if you install Cosmic Ray into and run workers from a virtual environment into which you've installed the modules under test.

Running distributed mutation testing

After you've started your workers, the only different between local and distributed tesing is that you need to pass --dist to the cosmic-ray exec command to do distributed testing. So a full distributed testing run would look something like this:

cosmic-ray init --baseline=3 session-name my_module -- tests
cosmic-ray exec --dist session-name
cosmic-ray report session-name


Cosmic Ray has a number of test suites to help ensure that it works. The first suite is a pytest test suite that validates some if its internals. You can run that like this:

py.test cosmic_ray/test

(Note that these unit tests don't require any workers to be running).

There is also a set of tests which verify the various mutation operators. These tests comprise a specially prepared body of code,, and a full-coverage test-suite. The idea here is that Cosmic Ray should be 100% lethal against the mutants of or there's a problem.

These tests can be run via both the standard unittest and py.test. In both cases, first make sure a worker (or several) is running. Then go to the test_project directory:

cd test_project

Run the operator tests with unittest like this:

cosmic-ray load cosmic-ray.unittest.conf

View the results of this test with report:

cosmic-ray report adam_tests.unittest

You should see a 0% survival rate at the end of the report.

Likewise you can run with py.test like this:

cosmic-ray load cosmic-ray.pytest.conf

The report will be available from the adam_tests.pytest session:

cosmic-ray report adam_tests.pytest


Mutation testing is conceptually simple and elegant. You make certain kinds of controlled changes (mutations) to your code, and then you run your test suite over this mutated code. If your test suite fails, then we say that your tests "killed" (i.e. detected) the mutant. If the changes cause your code to simply crash, then we say the mutant is "incompetent". If your test suite passes, however, we say that the mutant has "survived".

Needless to say, we want to kill all of the mutants.

The goal of mutation testing is to verify that your test suite is actually testing all of the parts of your code that it needs to, and that it is doing so in a meaningful way. If a mutant survives your test suite, this is an indication that your test suite is not adequately checking the code that was changed. This means that either a) you need more or better tests or b) you've got code which you don't need.

You can read more about mutation testing at the repository of all human knowledge. Lionel Brian has a nice set of slides introducing mutation testing as well.


Cosmic Ray works by parsing the module under test (MUT) and its submodules into abstract syntax trees using the ast module. It then uses the ast.NodeTransformer class to make systematic mutations to the ASTs.

For each individual mutation, Cosmic Ray modifies the Python runtime environment to replace the MUT with the mutated version. It then uses unittest's "discovery" functionality to discover your tests and run them against the mutant code.

In effect, the mutation testing algorithm is something like this:

for mod in modules_under_test:
    for op in mutation_operators:
        for site in mutation_sites(op, mod):
            mutant_ast = mutate_ast(op, mod, site)
            replace_module(, compile(mutant_ast)

                if discover_and_run_tests():
                    print('Oh no! The mutant survived!')
                    print('The mutant was killed.')
            except Exception:
                print('The mutant was incompetent.')

Obviously this can result in a lot of tests, and it can take some time if your test suite is large and/or slow.