[Python-ideas] Pre-conditions and post-conditions

[Marko Ristin-Kaufmann]

Hi David Maertz and Michael Lee,

Thank you for raising the points. Please let me respond to your comments in
separation. Please let me know if I missed or misunderstood anything.

*Assertions versus contracts.* David wrote:

> I'm afraid that in reading the examples provided it is difficulties for me
> not simply to think that EVERY SINGLE ONE of them would be FAR easier to
> read if it were an `assert` instead.
>

I think there are two misunderstandings on the role of the contracts.
First, they are part of the function signature, and not of the
implementation. In contrast, the assertions are part of the implementation
and are completely obscured in the signature. To see the contracts of a
function or a class written as assertions, you need to visually inspect the
implementation. The contracts are instead engraved in the signature and
immediately visible. For example, you can test the distinction by pressing
Ctrl + q in Pycharm.

Second, assertions are only suitable for preconditions. Postconditions are
practically unmaintainable as assertions as soon as you have multiple early
returns in a function. The invariants implemented as assertions are always
unmaintainable in practice (except for very, very small classes) -- you
need to inspect each function of the class and all their return statements
and manually add assertions for each invariant. Removing or changing
invariants manually is totally impractical in my view.

*Efficiency and Evidency. *David wrote:

> The API of the library is a bit noisy, but I think the obstacle it's more
> in the higher level design for me. Adding many layers of expensive runtime
> checks and many lines of code in order to assure simple predicates that a
> glance at the code or unit tests would do better seems wasteful.


I'm not very sure what you mean by expensive runtime checks -- every single
contract can be disabled at any point. Once a contract is disabled, there
is literally no runtime computational cost incurred. The complexity of a
contract during testing is also exactly the same as if you wrote it in the
unit test. There is a constant overhead due to the extra function call to
check the condition, but there's no more time complexity to it. The
overhead of an additional function call is negligible in most practical
test cases.

When you say "a glance at the code", this implies to me that you referring
to your own code and not to legacy code. In my experience, even simple
predicates are often not obvious to see in other people's code as one might
think (*e.g. *I had to struggle with even most simple ones like whether the
result ends in a newline or not -- often having to actually run the code to
check experimentally what happens with different inputs). Postconditions
prove very useful in such situations: they let us know that whenever a
function returns, the result must satisfy its postconditions. They are
formal and obvious to read in the function signature, and hence spare us
the need to parse the function's implementation or run it.

Contracts in the unit tests.

> The API of the library is a bit noisy, but I think the obstacle it's more
> in the higher level design for me. Adding many layers of expensive runtime
> checks and many lines of code in order to assure simple predicates that a
> glance at the code or *unit tests would do better* seems wasteful.
>
(emphasis mine)

Defining contracts in a unit test is, as I already mentioned in my previous
message, problematic due to two reasons. First, the contract resides in a
place far away from the function definition which might make it hard to
find and maintain. Second, defining the contract in the unit test makes it
impossible to put the contract in the production or test it in a call from
a different function. In contrast, introducing the contract as a decorator
works perfectly fine in all the three above-mentioned cases (smoke unit
test, production, deeper testing).

*Library. *Michael wrote:

> I just want to point out that you don't need permission from anybody to
> start a library. I think developing and popularizing a contracts library is
> a reasonable goal -- but that's something you can start doing at any time
> without waiting for consensus.


As a matter of fact, I already implemented the library which covers most of
the design-by-contract including the inheritance of the contracts. (The
only missing parts are retrieval of "old" values in postconditions and loop
invariants.) It's published on pypi as "icontract" package (the website is
https://github.com/Parquery/icontract/). I'd like to gauge the interest
before I/we even try to make a proposal to make it into the standard
library.

The discussions in this thread are an immense help for me to crystallize
the points that would need to be addressed explicitly in such a proposal.
If the proposal never comes about, it would at least flow into the
documentation of the library and help me identify and explain better the
important points.

*Observation of contracts. *Michael wrote:

> Your contracts are only checked when the function is evaluated, so you'd
> still need to write that unit test that confirms the function actually
> observes the contract. I don't think you necessarily get to reduce the
> number of tests you'd need to write.


Assuming that a contracts library is working correctly, there is no need to
test whether a contract is observed or not -- you assume it is. The same
applies to any testing library -- otherwise, you would have to test the
tester, and so on *ad infinitum.*

You still need to evaluate the function during testing, of course. But you
don't need to document the contracts in your tests nor check that the
postconditions are enforced -- you assume that they hold. For example, if
you introduce a postcondition that the result of a function ends in a
newline, there is no point of making a unit test, passing it some value and
then checking that the result value ends in a newline in the test.
Normally, it is sufficient to smoke-test the function. For example, you
write a smoke unit test that gives a range of inputs to the function by
using hypothesis library and let the postconditions be automatically
checked. You can view each postcondition as an additional test case in this
scenario -- but one that is also embedded in the function signature and
also applicable in production.

Not all tests can be written like this, of course. Dealing with a complex
function involves writing testing logic which is too complex to fit in
postconditions. Contracts are not a panacea, but they absolute us from
implementing trivial testing logic while keeping the important bits of the
documentation close to the function and allowing for deeper tests.

*Accurate contracts. *Michael wrote:

> There's also no guarantee that your contracts will necessarily be
> *accurate*. It's entirely possible that your preconditions/postconditions
> might hold for every test case you can think of, but end up failing when
> running in production due to some edge case that you missed.
>

Unfortunately, there is no practical exit from this dilemma -- and it
applies all the same for the tests. Who guarantees that the testing logic
of the unit tests are correct? Unless you can formally prove that the code
does what it should, there is no way around it. Whether you write contracts
in the tests or in the decorators, it makes no difference to accuracy.

If you missed to test an edge case, well, you missed it :). The
design-by-contract does not make the code bug-free, but makes the bugs *much
less likely* and *easier *to detect *early*. In practice, if there is a
complex contract, I encapsulate its complex parts in separate functions
(often with their own contracts), test these functions in separation and
then, once the tests pass and I'm confident about their correctness, put
them into contracts.

(And if you decide to disable those pre/post conditions to avoid the
> efficiency hit, you're back to square zero.)
>

In practice, we at Parquery AG let the critical contracts to run in
production to ensure that the program blows up before it exercises
undefined behavior in a critical situation. The informative violation
errors of the icontract library help us to trace the bugs more easily since
the relevant values are part of the error log.

However, if some of the contracts are too inefficient to check in
production, alas you have to turn them off and they can't be checked since
they are inefficient. This seems like a tautology to me -- could you please
clarify a bit what you meant? If a check is critical and inefficient at the
same time then your problem is unsolvable (or at least ill-defined);
contracts as well as any other approach can not solve it.

*Ergonimical assertions. *Michael wrote:

> Or I guess to put it another way -- it seems what all of these contract
> libraries are doing is basically adding syntax to try and make adding
> asserts in various places more ergonomic, and not much else. I agree those
> kinds of libraries can be useful, but I don't think they're necessarily
> useful enough to be part of the standard library or to be a technique
> Python programmers should automatically use by default.


>From the point of view of the *behavior, *that is exactly the case. The
contracts (*e.g. *as function decorators) make postconditions and
invariants possible in practice. As I already noted above, postconditions
are very hard and invariants almost impossible to maintain manually without
the contracts. This is even more so when contracts are inherited in a class
hierarchy.

Please do not underestimate another aspect of the contracts, namely the
value of contracts as verifiable documentation. Please note that the only
alternative that I observe in practice without design-by-contract is to
write contracts in docstrings in *natural language*. Most often, they are
just assumed, so the next programmer burns her fingers expecting the
contracts to hold when they actually differ from the class or function
description, but nobody bothered to update the docstrings (which is a
common pitfall in any code base over a longer period of time).

*Automatic generation of tests.* Michael wrote:

> What might be interesting is somebody wrote a library that does something
> more then just adding asserts. For example, one idea might be to try
> hooking up a contracts library to hypothesis (or any other library that
> does quickcheck-style testing). That might be a good way of partially
> addressing the problems up above -- you write out your invariants, and a
> testing library extracts that information and uses it to automatically
> synthesize interesting test cases.


This is the final goal and my main motivation to push for
design-by-contract in Python :). There is a whole research community that
tries to come up with automatic test generations, and contracts are of
great utility there. Mind that generating the tests based on contracts is
not trivial: hypothesis just picks elements for each input independently
which is a much easier problem. However, preconditions can define how the
arguments are *related*. Assume a function takes two numbers as arguments,
x and y. If the precondition is y < x < (y + x)  * 10, it is not trivial
even for this simple example to come up with concrete samples of x and y
unless you simply brute-force the problem by densely sampling all the
numbers and checking the precondition.

I see a chicken-and-egg problem here. If design-by-contract is not widely
adopted, there will also be fewer or no libraries for automatic test
generation. Honestly, I have absolutely no idea how you could approach
automatic generation of test cases without contracts (in one form or the
other). For example, how could you automatically mock a class without
knowing its invariants?

Since generating test cases for functions with non-trivial contracts is
hard (and involves collaboration of many people), I don't expect anybody to
start even thinking about it if the tool can only be applied to almost
anywhere due to lack of contracts. Formal proofs and static analysis are
even harder beasts to tame -- and I'd say the argument holds true for them
even more.

David and Michael, thank you again for your comments! I welcome very much
your opinion and any follow-ups as well as from other participants on this
mail list.

Cheers,
Marko

On Sat, 15 Sep 2018 at 10:42, Michael Lee <michael.lee.0x2a at gmail.com>
wrote:

> I just want to point out that you don't need permission from anybody to
> start a library. I think developing and popularizing a contracts library is
> a reasonable goal -- but that's something you can start doing at any time
> without waiting for consensus.
>
> And if it gets popular enough, maybe it'll be added to the standard
> library in some form. That's what happened with attrs, iirc -- it got
> fairly popular and demonstrated there was an unfilled niche, and so Python
> acquired dataclasses..
>
>
> The contracts make merely tests obsolete that test that the function or
>> class actually observes the contracts.
>>
>
> Is this actually the case? Your contracts are only checked when the
> function is evaluated, so you'd still need to write that unit test that
> confirms the function actually observes the contract. I don't think you
> necessarily get to reduce the number of tests you'd need to write.
>
>
> Please let me know what points *do not *convince you that Python needs
>> contracts
>>
>
> While I agree that contracts are a useful tool, I don't think they're
> going to be necessarily useful for *all* Python programmers. For example,
> contracts aren't particularly useful if you're writing fairly
> straightforward code with relatively simple invariants.
>
> I'm also not convinced that libraries where contracts are checked
> specifically *at runtime* actually give you that much added power and
> impact. For example, you still need to write a decent number of unit tests
> to make sure your contracts are being upheld (unless you plan on checking
> this by just deploying your code and letting it run, which seems
> suboptimal). There's also no guarantee that your contracts will necessarily
> be *accurate*. It's entirely possible that your
> preconditions/postconditions might hold for every test case you can think
> of, but end up failing when running in production due to some edge case
> that you missed. (And if you decide to disable those pre/post conditions to
> avoid the efficiency hit, you're back to square zero.)
>
> Or I guess to put it another way -- it seems what all of these contract
> libraries are doing is basically adding syntax to try and make adding
> asserts in various places more ergonomic, and not much else. I agree those
> kinds of libraries can be useful, but I don't think they're necessarily
> useful enough to be part of the standard library or to be a technique
> Python programmers should automatically use by default.
>
> What might be interesting is somebody wrote a library that does something
> more then just adding asserts. For example, one idea might be to try
> hooking up a contracts library to hypothesis (or any other library that
> does quickcheck-style testing). That might be a good way of partially
> addressing the problems up above -- you write out your invariants, and a
> testing library extracts that information and uses it to automatically
> synthesize interesting test cases.
>
> (And of course, what would be very cool is if the contracts could be
> verified statically like you can do in languages like dafny -- that way,
> you genuinely would be able to avoid writing many kinds of tests and could
> have confidence your contracts are upheld. But I understanding implementing
> such verifiers are extremely challenging and would probably have too-steep
> of a learning curve to be usable by most people anyways.)
>
> -- Michael
>
>
>
> On Fri, Sep 14, 2018 at 11:51 PM, Marko Ristin-Kaufmann <
> marko.ristin at gmail.com> wrote:
>
>> Hi,
>> Let me make a couple of practical examples from the work-in-progress (
>> https://github.com/Parquery/pypackagery, branch mristin/initial-version)
>> to illustrate again the usefulness of the contracts and why they are, in my
>> opinion, superior to assertions and unit tests.
>>
>> What follows is a list of function signatures decorated with contracts
>> from pypackagery library preceded by a human-readable description of the
>> contracts.
>>
>> The invariants tell us what format to expect from the related string
>> properties.
>>
>> @icontract.inv(lambda self: self.name.strip() == self.name)
>> @icontract.inv(lambda self: self.line.endswith("\n"))
>> class Requirement:
>>     """Represent a requirement in requirements.txt."""
>>
>>     def __init__(self, name: str, line: str) -> None:
>>         """
>>         Initialize.
>>
>>         :param name: package name
>>         :param line: line in the requirements.txt file
>>         """
>>         ...
>>
>> The postcondition tells us that the resulting map keys the values on
>> their name property.
>>
>> @icontract.post(lambda result: all(val.name == key for key, val in result.items()))
>> def parse_requirements(text: str, filename: str = '<unknown>') -> Mapping[str, Requirement]:
>>     """
>>     Parse requirements file and return package name -> package requirement as in requirements.txt
>>
>>     :param text: content of the ``requirements.txt``
>>     :param filename: where we got the ``requirements.txt`` from (URL or path)
>>     :return: name of the requirement (*i.e.* pip package) -> parsed requirement
>>     """
>>     ...
>>
>>
>> The postcondition ensures that the resulting list contains only unique
>> elements. Mind that if you returned a set, the order would have been lost.
>>
>> @icontract.post(lambda result: len(result) == len(set(result)), enabled=icontract.SLOW)
>> def missing_requirements(module_to_requirement: Mapping[str, str],
>>                          requirements: Mapping[str, Requirement]) -> List[str]:
>>     """
>>     List requirements from module_to_requirement missing in the ``requirements``.
>>
>>     :param module_to_requirement: parsed ``module_to_requiremnt.tsv``
>>     :param requirements: parsed ``requirements.txt``
>>     :return: list of requirement names
>>     """
>>     ...
>>
>> Here is a bit more complex example.
>> - The precondition A requires that all the supplied relative paths
>> (rel_paths) are indeed relative (as opposed to absolute).
>> - The postcondition B ensures that the initial set of paths (given in
>> rel_paths) is included in the results.
>> - The postcondition C ensures that the requirements in the results are
>> the subset of the given requirements.
>> - The precondition D requires that there are no missing requirements (*i.e.
>> *that each requirement in the given module_to_requirement is also
>> defined in the given requirements).
>>
>> @icontract.pre(lambda rel_paths: all(rel_pth.root == "" for rel_pth in rel_paths))  # A
>> @icontract.post(
>>     lambda rel_paths, result: all(pth in result.rel_paths for pth in rel_paths),
>>     enabled=icontract.SLOW,
>>     description="Initial relative paths included")  # B
>> @icontract.post(
>>     lambda requirements, result: all(req.name in requirements for req in result.requirements),
>>     enabled=icontract.SLOW)  # C
>> @icontract.pre(
>>     lambda requirements, module_to_requirement: missing_requirements(module_to_requirement, requirements) == [],
>>     enabled=icontract.SLOW)  # D
>> def collect_dependency_graph(root_dir: pathlib.Path, rel_paths: List[pathlib.Path],
>>                              requirements: Mapping[str, Requirement],
>>                              module_to_requirement: Mapping[str, str]) -> Package:
>>
>>     """
>>     Collect the dependency graph of the initial set of python files from the code base.
>>
>>     :param root_dir: root directory of the codebase such as "/home/marko/workspace/pqry/production/src/py"
>>     :param rel_paths: initial set of python files that we want to package. These paths are relative to root_dir.
>>     :param requirements: requirements of the whole code base, mapped by package name
>>     :param module_to_requirement: module to requirement correspondence of the whole code base
>>     :return: resolved depedendency graph including the given initial relative paths,
>>     """
>>
>> I hope these examples convince you (at least a little bit :-)) that
>> contracts are easier and clearer to write than asserts. As noted before in
>> this thread, you can have the same *behavior* with asserts as long as
>> you don't need to inherit the contracts. But the contract decorators make
>> it very explicit what conditions should hold *without* having to look
>> into the implementation. Moreover, it is very hard to ensure the
>> postconditions with asserts as soon as you have a complex control flow since
>> you would need to duplicate the assert at every return statement. (You
>> could implement a context manager that ensures the postconditions, but a
>> context manager is not more readable than decorators and you have to
>> duplicate them as documentation in the docstring).
>>
>> In my view, contracts are also superior to many kinds of tests. As the
>> contracts are *always* enforced, they also enforce the correctness
>> throughout the program execution whereas the unit tests and doctests only
>> cover a list of selected cases. Furthermore, writing the contracts in these
>> examples as doctests or unit tests would escape the attention of most less
>> experienced programmers which are not  used to read unit tests as
>> documentation. Finally, these unit tests would be much harder to read than
>> the decorators (*e.g.*, the unit test would supply invalid arguments and
>> then check for ValueError which is already a much more convoluted piece of
>> code than the preconditions and postconditions as decorators. Such testing
>> code also lives in a file separate from the original implementation making
>> it much harder to locate and maintain).
>>
>> Mind that the contracts *do not* *replace* the unit tests or the
>> doctests. The contracts make merely tests obsolete that test that the
>> function or class actually observes the contracts. Design-by-contract helps
>> you skip those tests and focus on the more complex ones that test the
>> behavior. Another positive effect of the contracts is that they make your
>> tests deeper: if you specified the contracts throughout the code base, a
>> test of a function that calls other functions in its implementation will
>> also make sure that all the contracts of that other functions hold. This
>> can be difficult to implement  with standard unit test frameworks.
>>
>> Another aspect of the design-by-contract, which is IMO ignored quite
>> often, is the educational one. Contracts force the programmer to actually
>> sit down and think *formally* about the inputs and the outputs
>> (hopefully?) *before* she starts to implement a function. Since many
>> schools use Python to teach programming (especially at high school level),
>> I imagine writing contracts of a function to be a very good exercise in
>> formal thinking for the students.
>>
>> Please let me know what points *do not *convince you that Python needs
>> contracts (in whatever form -- be it as a standard library, be it as a
>> language construct, be it as a widely adopted and collectively maintained
>> third-party library). I would be very glad to address these points in my
>> next message(s).
>>
>> Cheers,
>> Marko
>>
>> _______________________________________________
>> Python-ideas mailing list
>> Python-ideas at python.org
>> https://mail.python.org/mailman/listinfo/python-ideas
>> Code of Conduct: http://python.org/psf/codeofconduct/
>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/python-ideas/attachments/20180915/b808a93c/attachment-0001.html>