Nice concept but does it work with real-world application.
I failed to understand how will it work with methods like `authenticate_user(user)` or `load_permissions_from_db(user, db)`.
However, I would also note, that just because you have some methods that have potentially dangerous side effects, in most large code-bases, not all functions do. Which is why it operates at a function by function basis. You can use this for new pure functions, while continuing to write other tests in the same fashion by hand.
Thanks for the write up - explains a lot (including what hypothesis is).
I think there's a typo in the last code box? It looks like you repeated "from hypothesis_auto import"? (that, or I understand even less python that I thought!)
Does this do anything other than assert that functions return the right basic type given extremely basic inputs?
Is naive type testing a thing that people actually bother doing vs testing that the functions actually do what they're supposed to?
If a function takes two ints and returns an int that is the integer division of the two, is the evaluation of exceptions and returned types not already implicit in the evaluation of the actual results?
If you know that divide(a, b) should return c for sufficient candidates a, b, and c, then you _know_ that divide returns the right type without explicitly checking. And knowing that the divide function happens to return ints when given ints doesn't actually tell you that it's doing anything even close to the right behavior. So this both doesn't reduce the number of tests you need to write and is also obsoleted by actually writing the tests that you need.
By default, the biggest value from this is catching unexpected runtime exceptions with certain values (the edge cases you don't think of yourself). To get the most value out of it, you can further specify what you expect from the result using the _auto_verify=Callable() parameter, or by looking at the individual test cases.
Simple example:
from my_library import add
from hypothesis_auto import auto_pytest
@auto_pytest()
def test_add(test_case):
add_result = test_case()
if test_case.params.kwargs['number_1'] > 0 and test_case.params.kwargs['number_1']:
assert add_result > test_case.params.kwargs['number_1']
assert add_result > test_case.params.kwargs['number_1']
It's also useful to check the function doesn't crash.
I use hypothesis to test a DB abstraction, and it often caught bugs in edge cases I didn't consider: timestamps/timezones too large, strings containing null bytes, etc.
These would be caught by the tests generated by hypothesis-auto.
It's an extension for hypothesis, and its value is 100% just convenience and accessibility. Really, it's meant to be a sort of gateway to doing full property-based testing, in the least barrier way possible. I wrote a bit more about why I created it here: https://timothycrosley.com/project-5-hypothesis-auto
Hypothesis' `infer` is inferring arguments to the test fn, that you'd then use to drive the fn under test.
OP's extension is inferring arguments to the fn under test, and then generating a test like 'assert isinstancee(fn_under_test(a, *kw), expected_ret_type)` to go with them.
You can do more types of test with `infer` (i.e. not just 'does return' and 'returns correct type' but also 'returns correct value'), but this is an easy way to cover a lot of basic return checking ground.
I like this idea. One piece of feedback, a parameter with a leading underscore feels very odd. In python I interpret leading underscores to indicate the programmer thinks of this as an internal / pseudo-private property. Exposing it through the api makes it "public" which means (to me) that it shouldn't have a leading underscore.
This is especially true if the usecase is common enough to put in the top level examples
While I agree, that `_param` generally means private, it is also a common way to allow parameters to be passed into a function where all other parameters are passed directly along via args and *kwargs to avoid naming collisions. An example of this is NamedTuple: https://docs.python.org/3/library/collections.html#namedtupl....
I believe Raymond Hettinger now considers that a mistake, and wishes he had gone with a trailing underscore (param_) instead. A trailing underscore is just as unlikely to lead to a collision, but less confusing.
It takes a bit of manual juggling, but if you take kwargs anyway you could inspect them to check for the old-style parameters, and use them if they're present, perhaps with a warning. That way you don't have to break compatibility.
This actually gave me another idea. What do you all think, I’d be up for trying to build it.
It would watch your program during execution and record each function calls input and output. And then create tests for each function using those inputs and outputs.
You could always go over the created tests manually and fix them or review them but if nothing else it could be a good start.
I really like this idea. So, broken out, would the development process look like:
1) Write code
2) Open interpreter (ipython)
3) Import code and call it manually
4) This magic library will have observed and created a testcase module from your execution
If the library can use the observed and recorded testcase to extrapolate far beyond your manual call, this could be a really interesting way of testing python code.
This is actually the idea behind a little library I developed for my own use in common lisp called testy.
Because the common lisp environment already defines special variables that contain the last form evaluated and the return value of the last form evaluated, when you evaluate code at the REPL and it succeeds, you call the (testy:test) function, which captures them and wraps them up as a test for the package you're working on, adding it to the database of tests for the package.
Not only was it useful for development (no need to specify your tests up front, you write your code until it works, then that state "is" the test), but it was also helpful for fixing bugs that you hadn't thought of up front (you fix the bug and then that state is captured as a new test as well).
Instagram created 'MonkeyType' to do something similar for type annotations.
A lower barrier way of trying this out could be to run MonkeyType in combination with OP's hypothesis-auto (which uses type annotations to generate the tests).
While I had really liked the idea of hypothesis in Python I found that the edge-cases it was uncovering were the ones that were obviously gonna break but at the same time cases I didn't care to guard against, e.g., using 3-mile long integers, or cases that wouldn't work with the underlying libraries eg NumPy. Thus, I found myself spending more time adding constraints on the generated inputs than fledging out my test-suite. So my adventures with hypothesis were short-lived.
I don't mean to detract from this library, I think its a great combination of strong-typing and property-based testing but has anyone had any experience employing property-based testing on complex functions outside of the whole add/subtract/multiple stuff? What kinda thing have you used it on?
38 comments
[ 3.2 ms ] story [ 77.1 ms ] threadFor cases like that it would be best combined with a mock:
Note that this example utilizes the following pytest extension: https://github.com/pytest-dev/pytest-mockHowever, I would also note, that just because you have some methods that have potentially dangerous side effects, in most large code-bases, not all functions do. Which is why it operates at a function by function basis. You can use this for new pure functions, while continuing to write other tests in the same fashion by hand.
I think there's a typo in the last code box? It looks like you repeated "from hypothesis_auto import"? (that, or I understand even less python that I thought!)
Is naive type testing a thing that people actually bother doing vs testing that the functions actually do what they're supposed to?
If a function takes two ints and returns an int that is the integer division of the two, is the evaluation of exceptions and returned types not already implicit in the evaluation of the actual results?
If you know that divide(a, b) should return c for sufficient candidates a, b, and c, then you _know_ that divide returns the right type without explicitly checking. And knowing that the divide function happens to return ints when given ints doesn't actually tell you that it's doing anything even close to the right behavior. So this both doesn't reduce the number of tests you need to write and is also obsoleted by actually writing the tests that you need.
Simple example:
So it's a fuzzer?
I use hypothesis to test a DB abstraction, and it often caught bugs in edge cases I didn't consider: timestamps/timezones too large, strings containing null bytes, etc.
These would be caught by the tests generated by hypothesis-auto.
OP's extension is inferring arguments to the fn under test, and then generating a test like 'assert isinstancee(fn_under_test(a, *kw), expected_ret_type)` to go with them.
You can do more types of test with `infer` (i.e. not just 'does return' and 'returns correct type' but also 'returns correct value'), but this is an easy way to cover a lot of basic return checking ground.
This is especially true if the usecase is common enough to put in the top level examples
It takes a bit of manual juggling, but if you take kwargs anyway you could inspect them to check for the old-style parameters, and use them if they're present, perhaps with a warning. That way you don't have to break compatibility.
Also, `auto_pytest_magic` doesn't seem to exist.
This is a very interesting idea I might explore more!
`auto_pytest_magic` is only made available if you have pytest installed. You can enforce this by doing:
It would watch your program during execution and record each function calls input and output. And then create tests for each function using those inputs and outputs.
You could always go over the created tests manually and fix them or review them but if nothing else it could be a good start.
If the library can use the observed and recorded testcase to extrapolate far beyond your manual call, this could be a really interesting way of testing python code.
Because the common lisp environment already defines special variables that contain the last form evaluated and the return value of the last form evaluated, when you evaluate code at the REPL and it succeeds, you call the (testy:test) function, which captures them and wraps them up as a test for the package you're working on, adding it to the database of tests for the package.
Not only was it useful for development (no need to specify your tests up front, you write your code until it works, then that state "is" the test), but it was also helpful for fixing bugs that you hadn't thought of up front (you fix the bug and then that state is captured as a new test as well).
[0] https://github.com/vcr/vcr
A lower barrier way of trying this out could be to run MonkeyType in combination with OP's hypothesis-auto (which uses type annotations to generate the tests).
[1] https://github.com/google/pytype
[2] https://github.com/dropbox/pyannotate
[3] https://github.com/Instagram/MonkeyType
[4] https://news.ycombinator.com/item?id=19454411
I know, coverage is Yet Another Metric, but if you don't game it, it can help you track down branches you haven't written tests for.
So my hesitation is I can see people running this, gettting 100% code coverage and thinking, "hooray, it's fully tested!"
I don't mean to detract from this library, I think its a great combination of strong-typing and property-based testing but has anyone had any experience employing property-based testing on complex functions outside of the whole add/subtract/multiple stuff? What kinda thing have you used it on?