ARCHITECTURE.md

May 1, 2026 ยท View on GitHub

Choices

There are a number of choices when writing a Python type checker. We are taking inspiration from Pyre1, Pyright and MyPy. Some notable choices:

  • We infer types in most locations, apart from parameters to functions. We do infer types of variables and return types. As an example, def foo(x): return True would result in something equivalent to had you written def foo(x: Any) -> bool: ....
  • We attempt to infer the type of [] to however it is used first, then fix it after. For example xs = []; xs.append(1); xs.append("") will infer that xs: List[int] and then error on the final statement.
  • We use flow types which refine static types, e.g. x: int = 4 will both know that x has type int, but also that the immediately next usage of x will be aware the type is Literal[4].
  • We aim for large-scale incrementality (at the module level) and optimized checking with parallelism, aiming to use the advantages of Rust to keep the code a bit simpler.
  • We expect large strongly connected components of modules, and do not attempt to take advantage of a DAG-shape in the source code.

Code layout

Pyrefly is split into a number of crates (mostly under crates/):

  • pyrefly_util are general purpose utilities, which have nothing to do with Python or type checking. Examples include IO wrappers, locking, command line helpers etc.
  • pyrefly_derive are proc-macros for deriving traits such as TypeEq and Visit.
  • pyrefly_python are Python utilities with no type-checking aspects, such as modelling modules or sys.info.
  • pyrefly_graph provides utilities for indexing values and caching computations that may depend on each other.
  • pyrefly_bundled are the third-party typeshed stubs.
  • pyrefly_config defines the Pyrefly configuration, along with support for reading Mypy/Pyright configuration.
  • pyrefly_types defines the Pyrefly type along with operations on it.
  • pyrefly_wasm defines the sandbox code that compiles to WASM.
  • pyrefly itself is the type checker and everything else.

Design

There are many nuances of design that change on a regular basis. But the basic substrate on which the checker is built involves three steps:

  1. Figure out what each module exports. That requires solving all import * statements transitively.
  2. For each module in isolation, convert it to bindings, dealing with all statements and scope information (both static and flow).
  3. Solve those bindings, which may require the solutions of bindings in other modules.

If we encounter unknowable information (e.g. recursion) we use Type::Var to insert placeholders which are filled in later.

For each module, we solve the steps sequentially and completely. In particular, we do not try and solve a specific identifier first (like Roslyn or TypeScript), and do not use fine-grained incrementality (like Rust Analyzer using Salsa). Instead, we aim for raw performance and a simpler module-centric design - there's no need to solve a single binding in isolation if solving all bindings in a module is fast enough.

Example of bindings

Given the program:

1: x: int = 4
2: print(x)

We might produce the bindings:

  • define int@0 = from builtins import int
  • define x@1 = 4: int@0
  • use x@2 = x@1
  • anon @2 = print(x@2)
  • export x = x@2

Of note:

  • The keys are things like define (the definition of something), use (a usage of a thing) and anon (a statement we need to type check, but don't care about the result of).
  • In many cases the value of a key refers to other keys.
  • Some keys are imported from other modules, via export keys and import values.
  • In order to disambiguate identifiers we use the textual position at which they occur (in the example we've used @line, but in reality it's the byte offset in the file).

Example of Var

Given the program:

1: x = 1
2: while test():
3:     x = x
4: print(x)

We end up with the bindings:

  • x@1 = 1
  • x@3 = phi(x@1, x@3)
  • x@4 = phi(x@1, x@3)

The expression phi is the join point of the two values, e.g. phi(int, str) would be int | str. We skip the distinction between define and use, since it is not necessary for this example.

When solving x@3 we encounter recursion. Operationally:

  • We start solving x@3.
  • That requires us to solve x@1.
  • We solve x@1 to be Literal[1]
  • We start solving x@3. But we are currently solving x@3, so we invent a fresh Var (let's call it ?1) and return that.
  • We conclude that x@3 must be Literal[1] | ?1.
  • Since ?1 was introduced by x@3 we record that ?1 = Literal[1] | ?1. We can take the upper reachable bound of that and conclude that ?1 = Literal[1].
  • We simplify x@3 to just Literal[1].