Development Hints
Deterministic behavior
Pynguin’s behavior is inherently probabilistic. However, in order to easily reproduce an execution, e.g. for an evaluation or for debugging purposes, it must be possible to achieve deterministic behavior.
For this purpose, whenever an implementation uses randomness, it must be done via pynguin.utils.randomness
.
This module contains a singleton instance of a (pseudo) random-number generator that is seeded at startup (using --seed
).
Furthermore, when an implementation’s behavior depends on the iteration order of a data structure, e.g., when picking a random element from it, one has to ensure that this order is deterministic:
lists and tuples are ordered by design.
dicts store their insertion order beginning with Python 3.7 and are therefore safe to use.
Python’s builtin sets do not guarantee any order, thus one has to use
OrderedSet
, which is a set implementation that stores the insertion order of its elements.
Overriding __eq__
and __hash__
methods
Similar to the Java world, we enforce to adhere to contracts for equals and hash
methods as described, for example, in the Java API documentation.
More information can also be found in Joshua Bloch’s famous book Effective Java.
In particular, when you override the __eq__
or the __hash__
method of a class
you are also required to override its opponent.
The following contracts should hold (adopted from the Java API documentation):
For
__hash__
__hash__
must consistently return the same integer whenever it is invoked on the same object more than once during one execution of the Python application, provided no information used in__eq__
comparisons on the same object is modified. The integer need not remain consistent from one application execution to another execution of the same application.If two objects are equal according to the
__eq__
method then the__hash__
value of the two objects must be the sameIt is not required that for two objects being unequal according to the
__eq__
method, the__hash__
value must be distinct, although this could improve performance of hash tables.
For
__eq__
The relation is reflexive: for any non-null reference value
x
,x.__eq__(x)
should returnTrue
The relation is symmetric: for any non-null reference values
x
andy
,x.__eq__(y)
should returnTrue
if and only ify.__eq__(x)
returnsTrue
.The relation is transitive: for any non-null reference values
x
,y
, andz
, ifx.__eq__(y)
returnsTrue
andy.__eq__(z)
returnsTrue
, then alsox.__eq__(z)
should returnTrue
.The relation is consistent: Multiple invocations of the method on the same two objects should yield the same result as long as none of the objects has been changed.
For any non-null reference value
x
,x.__eq__(None)
should returnFalse
Overriding __str__
and __repr__
methods
The goal of a __str__
method is to provide a string representation that is
usable and readable for a user. The goal of the __repr__
method is to be
unambiguous, see StackOverflow.
We encourage you to provide a __repr__
representation that looks like the Python
code that creates an object with the state of the object __repr__
was called on.
Consider the following example:
1class Example:
2 def __init__(
3 self, foo: str, bar: int, baz: List[str]
4 ) -> None:
5 self._foo = foo
6 self._bar = bar
7 self._baz = baz
8
9
10example = Example("abc", 42, ["xyz", "pynguin"])
The representation, i.e., the result yielded by calling __repr__
on the
example
object should look like
Example(foo="abc", bar=42, baz=["xyz", "pynguin"])
which can be achieved by implementing the __repr__
method of the Example
class as follows:
1 def __repr__(self) -> str:
2 return f"Example(foo=\"{self._foo}\", bar={self._bar}, "
3 f"baz={repr(self._baz)})"
Guarding imports for type checking
Some imports in a module are only necessary for type checking but not at runtime.
We guard these imports by if typing.TYPE_CHECKING
blocks.
The main reason for this is to prevent circular imports.
During type checking, these imports do not harm the type checker as it uses much more
sophisticated techniques to handle the circular imports (like a compiler does) in
contrast to the simple handling of the interpreter.
Debugging test case execution
We execute test cases in a separate thread.
To track data on the test case execution, e.g., line or branch coverage, we use thread-local storage.
Usage of threading.local may interfere with debugging tools, such as pydevd.
In such a case, disable Cython by setting the following environment variable:
PYDEVD_USE_CYTHON=NO
Comments and DocStrings
We have no general policy regarding comments in the source code. Use them, whenever you feel they are necessary. Please do not explain what the code is doing, but why.
DocStrings are required for all public functions, methods, constructors, classes, and modules. We use the
ruff
linter to check for the DocStrings. You can omit the DocString for default constructors, i.e., constructors that take no arguments, and methods that override a method from a parent class. In the former case, disableruff
’s warning by adding# noqa: D107
to the line of the constructor declaration; use# noqa: D102
in the latter case, respectively.Please follow the Google style for your DocString formatting.
Sometimes, the interface of a class forces you to override a method, e.g., because the base class is abstract, but there is no need for a concrete implementation. In such a case put the following DocString to the overriding method to show that it is empty on purpose.