Frequently asked questions¶
Why another typechecker?¶
Pyanalyze started at Quora as an internal tool somewhere between a linter and a type checker, but it proved very useful in dealing with asynq, our asynchronous programming framework. This framework uses generators in an unusual way and pyanalyze made it possible to detect several tricky mistakes in asynq code statically.
For us, asynq support remains important, but pyanalyze’s architecture has also allowed us to perform numerous other static checks to make our codebase safer. For example, we use pyanalyze to help keep our codebase safe against SQL injections, to enforce that names of A/B tests are valid, and to enforce that UI strings are translated correctly.
What makes pyanalyze different?¶
The biggest architectural difference between pyanalyze and other Python type checkers is that pyanalyze imports the code it checks, while other checkers purely look at the source code. This allows features that are very difficult to achieve with a fully static type checker:
pyanalyze requires no special casing to understand the semantics of the
@dataclass
decorator, which creates a synthesized__init__
method. It simply inspects the signature of the generated method and uses that for type checking. In general, many dynamic constructs unsupported by other type checkers will work immediately with pyanalyze.pyanalyze can call back into user code to customize type checking behavior. For example, the
CustomCheck
extension provides a way for user code to get very precise control over type checking behavior. Possible use cases include allowing only literal values, disallowing usage ofAny
for specific APIs, and allowing only values that can be pickled at runtime.
But pyanalyze is still a static checker, and it has some advantages over a dynamic (runtime) typechecker:
All code paths are checked, not just the ones that are hit in a particular run.
The type system can carry around more information than just the runtime type of a value. For example, pyanalyze supports
NewType
wrappers around runtime types.pyanalyze can use type stubs such as those in typeshed for type checking.
In addition, pyanalyze checks each module mostly independently, keeping the AST for only one module in memory at once and using runtime function and module objects for computing signatures and types. This reduces memory usage and makes it easier to deal with circular dependencies.
However, this approach also has some disadvantages:
It is difficult to engineer an incremental mode, because that would require reloading imported modules. At Quora, we run pyanalyze locally on changed files only, which is much faster than running it on the entire codebase but does not catch issues where a change breaks type checking in an upstream module.
Scripts that do work on import cannot be usefully checked; you must guard execution with
if __name__ == "__main__":
.Undefined attributes on instances of user-defined classes cannot be detected with full confidence, because the class object does not provide a good way to find out which attributes are created in the
__init__
method. Currently, pyanalyze works around this by deferring detection of undefined attributes until the entire codebase has been checked, but this is fragile and not always reliable.Initially, the implementation of
@typing.overload
did not provide a way to access the overloads at runtime, so there was no obvious way to support overloaded functions at runtime. However, this was fixed in Python 3.11.
When should I use pyanalyze?¶
If you have a complex Python codebase and want to make sure it stays maintainable and stable, using a type checker is a great option. There are several options for type checking Python code (listed in the typing documentation). Unique advantages of pyanalyze include:
Better support for dynamic constructs and configurability thanks to its semi-static architecture (see “What makes pyanalyze different?” above).
Support for specific checks, such as finding missing f prefixes in f-strings, finding missing
await
statements, detecting possibly undefined names, and warning about conditions on objects of suspicious types.Type system extensions such as
CustomCheck
,ParameterTypeGuard
, andExternalType
.Strong support for checking code that uses the asynq framework.
What is the history of pyanalyze?¶
The first incarnation of pyanalyze dates back to July 2015 and merely detected undefined names. It was soon extended to detect incompatible function calls, find unused code, and perform some basic type checking.
For context, pytype was started in March 2015 and mypy in 2012. PEP 484 was accepted in 2015.
The initial version was closely tied to
Quora’s internal code structure, but in June 2017 it was split off into its own internal
package, now named pyanalyze. By then, it had strong support for asynq
and supported
customization through implementation functions. After Quora moved to Python 3, pyanalyze
gained support for parsing type annotations and
its typing support moved closer to the standard type system.
The first public release was in May 2020. Since then, work has focused on providing full
support for the Python type system, including Annotated
, Callable
, TypeVar
, and
TypeGuard
.