rattr 0.1.0

Rattr rats on your attrs.

Rattr rats on your attrs.

Rattr (pronounced 'ratter') is a tool to determine attribute usage in python functions. It can parse python files, follow imports and then report to you about the attributes accessed by function calls in that file.

Status

Currently this project is under active development and likely to change significantly in the future. However we hope it might be useful and interesting to the wider python community.

But why?

We developed rattr to help with some analytics work in python where type checkers like mypy and pyre are cumbersome. In analytics work, we often have functions that look like this:

def compute_cost_effectiveness(person):
    return person.sales / person.salary

because we're pythonistas, the exact type of person is unimportant to us in this example - what's important is that it has a sales and salary attribute and that those are numbers. Annotating this function with that information for mypy would be cumbersome - and with thousands of functions it would be hard to do.

Rattr is a tool that solves the first part of this - it can detect that compute_cost_effectiveness needs to access "sales" and "salary" attributes and so it could tell us that the following would fail:

def create_report():
    people = some_database.query(Person.name, Person.sales).all()
    return {person.name: compute_cost_effectiveness(person) for person in people}

It can also effectively compute the provenance of attributes. Suppose that you have a wide array of functions for computing information about financial products - like

def compute_some_complex_risk_metric_for(security, other_data):
    # proprietary and complicated logic here
    security.riskiness = bla
    return security

and you have other functions that consume that information:

def should_i_buy(security):
    if security.riskiness > 5:
        return False
    # More logic here ...

rattr can help you determine which functions are required for a calculation. Effectively allowing you to build powerful directed graph structures for your function libraries.

Developer Notes

Use of Undocumented Behaviour

In rattr/analyser/types.py several Union types are defined for convenience. In Python 3.8 to check if the variable a is an instance of any of the types within the Union the typing module provides get_args i.e. one would use isinstance(a, get_args(UnionTypeName)). However, this function is not provided in Python 3.7 and so the undocumented attribute __args__ of the UnionTypeName must be used i.e. isinstance(a, UnionTypeName.__args__). As this is undocumented it should be changed when we upgrade to Python 3.8+ (and $EDITOR will not syntax highlight or tab-complete it).

Annotations

Rattr provides the ability to annotate functions in the target file such that they may be ignored completely, ignored with their results determined manually, etc. Additionally, each assertor may provide it's own annotations to ignore and/or define behaviour.

General Rattr annotations are located in rattr/analyser/annotations.py and assertor specific annotations are to be added by the assertion developer -- however they should be placed in the file containing the Assertor class.

Annotation Format

Annotations should take the form rattr_<annotation_name> to avoid namespace conflicts in importing code.

Detecting and Parsing Annotations

The rattr/analyser/utils.py file provides the following annotation utility functions:

• has_annotation(name: str, fn: ast.FunctionDef) -> bool
• get_annotation(name: str, fn: ast.FunctionDef) -> Optional[ast.AST]
• parse_annotation(name: str, fn: ast.FunctionDef) -> Dict[str, Any]
• parse_rattr_results_from_annotation(fn: ast.FunctionDef) -> Dict[str, Literal[...]]:
• safe_eval(expr: ast.expr, culprit: Optional[ast.AST]) -> Union[Literal, Iterable[Iterable[...[Literal]]]]
• is_name(name: Any) -> bool
• is_set_of_names(set_of_names: Any) -> bool
• is_args(args: Any) -> bool

Provided Annotations

Annotation Name	Location
rattr_ignore	rattr/analyser/annotations.py
rattr_results(<results>)	rattr/analyser/annotations.py

Results Annotation Grammar

@rattr_results(
    sets={"a", "b.attr"},
    calls=[
        ("callee_function", (["arg", "another"], {"kwarg": "some_var"}))
    ]
)
def decorated_function(...):
    # ...

Any argument to the decorator can be omitted and a default value will be used.

Known Issues

Nested functions are not currently analysed properly, functions containing nested functions must be annotated manually.

Comprehensions are not fully analysed, should be solvable by the same approach as nested functions -- "un-nest" them.

---

Usage Notes

See python rattr -h.

Errors and Warnings

Rattr can give five types of error/warnings: raised Python errors, output beginning with "info:" or "warning:", output beginning with "error:", and output beginning with "fatal:". The former can be seen as a developer caused error, and the latter four are user errors.

User Error: "info" and "warning"

Warns the user of potential issues or bad practise that should not affect the results of analysis. Low-priority (often class based) warnings begin with "info".

User Error: "error"

Warns the user of potential issues or bad practise that will likely affect the results of analysis (though there are times when the results will still be correct).

User Error: "fatal"

Warns the user of potential issues or bad practise so severe that the results can not be produced for the given file.

Results Structure

A dictionary from functions to results, which is in turn a dictionary of the variables, attributes, etc (collectively nameables) that are get, set, called, or deleted.

Nameables Format

Pythonic Name	Python Example	Rattr Result Format
Name	name	name
Starred	*name	*name
Attribute	name.attr	name.attr
Call	name(a, b, ...)	name()
Subscript	name[0] or name['key']	name[]

The above can be nested in Rattr as in Python, for example the Python snippet name.method(arg_one, arg_two, ...).result_attr['some key'] will become name.method().result_attr[].

However, some expression do not have resolvable names. For example, given the class A and instances a_one, a_two; assuming that A implements __add__, which over two names of type A returns an A; and, A provides the attribute some_attr; the following is legal Python code (a_one + a_two).some_attr. Another example of code whose name is unresolvable is (3).to_bytes(length=1, byteorder='big').

Rattr will handle the above cases by returning a produced local name -- the result of prepending the AST node type with an '@'. The former example will become @BinOp.some_attr, and the latter @Int.to_bytes.

Example Results

{
    ...

    "my_function": {
        "gets": [
            "variable_a",
            "variable_b",
            "object_a.some_attr",
        ],
        "sets": [
            "object_a.set_me",
        ],
        "dels": [],
        "calls": [
            "min()",
            "max()"
        ]
    },
}

Support for Python 3.8

Between Python 3.7 and Python 3.8 there were several significant changes that effect Rattr and how it works, namely:

1. the introduction of the walrus operator;
2. the addition of posonlyargs to ast.arguments;
3. complete rework of the representation of constants in ast.

As it stands Rattr will run on-and-under Python 3.8, however, with varying support for the above. Specifically: 1. is not supported, and usage will cause an error; 2. is not supported, and usage results in undefined behaviour; and 3. is fully supported.

An additional issue with Python 3.7/3.8 cross-compatibility is the introduction of typing.get_origin and typing.get_args, removing the need to rely on the undefined behaviour of Union().__args__. Though, as the latter works in both, this does not affect the execution of Rattr -- it just results in some complaints by mypy which could be avoided it only Python 3.8 were supported.

Links regarding the above:

[1] https://stackoverflow.com/questions/45957615/check-a-variable-against-union-type-at-runtime-in-python-3-6 -- answers 1 and 2 specifically touch on the typing.get_args / Union().__args__ issue.

[2] https://greentreesnakes.readthedocs.io/en/latest/nodes.html#literals -- literals are very different.

[3] https://greentreesnakes.readthedocs.io/en/latest/nodes.html#NamedExpr -- walrus operator.

[4] https://greentreesnakes.readthedocs.io/en/latest/nodes.html#arguments -- posonlyargs.

---

Known Issues

For now these will throw a fatal error -- in the future Rattr should be more "feature complete" and should handle these cases properly.

Globals

>>> y = 8
>>> def fn(a):
...     global y
...     y = a
...
>>> fn(3)
>>> y
3
>>>

Redefinitions

def fn(m):
    print(m.attr_one)

    m = MyCoolClass()

    print(m.attr_two)

Rattr will produce { "sets": {"m.attr_one", "m.attr_two"} }
But should produce { "sets": {"m.attr_one", "@local:m.atter_two"} }?