rustruct

A Rust core for parsing and building binary wire formats from Python.

A declarative schema built from Python primitives is compiled into a program executed with a single Python -> Rust call for the whole structure. The parse result is a plain dict.

Building and testing

# Rust core (no Python)
cargo test -p rustruct-core

# Python package (maturin backend) + pytest
uv sync
uv run pytest

# Sphinx documentation (warnings are errors)
make test-docs

Example

from rustruct import Field, Incomplete, compile

codec = compile(
    (
        Field(name="tag", kind="u8", opts={}),
        Field(name="size", kind="u8", opts={}),
        Field(
            name="value",
            kind="struct",
            opts={
                "fields": (Field(name="payload", kind="bytes", opts={"len": "*"}),),
                "size": ("ref", "size"),
            },
        ),
        Field(name="crc", kind="digest", opts={"algo": "crc32", "over": "*"}),
    )
)

data = codec.pack({"tag": 7, "value": {"payload": b"abc"}})
# size and crc are derived: computed and patched in automatically

values = codec.unpack(data)  # full buffer, a tail is an error
values, pos = codec.unpack_from(data, 0)  # a tail is allowed

r = codec.parse(data[:3])  # streaming parse
if not r:  # Incomplete: not enough data yet
    print("need at least", r.needed, "more bytes")

Documentation

The full documentation is at mosquito.github.io/rustruct and is organised by user need: a guided tutorial, task-oriented how-to guides, explanations of the schema and execution model, and concise API reference.

Declarative frontend

A Struct metaclass on top of compile()/Codec: class-body annotations and field descriptors (described, slice, array, switch, bits, convert, sized, an open registry()) compile lazily, per class, into a Codec, and convert to/from typed instances rather than plain dicts.

from rustruct import Struct, U8, U16, switch, slice, registry

Payload = Struct  # a registry base: class Foo(Payload, registry=True): ...

See src/rustruct/protocols/{inet,tcp,udp,dns}.py for a full worked example (IPv4/TCP/UDP header dispatch via an open registry, DNS with hand-written name-compression as the one piece that doesn't fit the declarative model) and tests/test_frontend_*.py, tests/test_protocols_*.py for the tests.

Repository layout

crates/rustruct/       core (rustruct-core): schema compiler, unpack/pack,
                       expressions, bits, flags, windows, digests; zero pyo3
crates/rustruct-py/    pyo3 cdylib -> the rustruct.core module
src/rustruct/          Python wrapper: low-level re-export, __abi__ check,
                       the Struct frontend (struct.py/fields.py/scalars.py/
                       expr.py), and protocols/ (IPv4/TCP/UDP/DNS example)
tests/                 pytest tests: public API, frontend, protocols
docs/                  tutorials, how-to guides, explanation, API reference
crates/rustruct/tests/ Rust integration tests, one file per feature, plus
                       tests/common/mod.rs for shared helpers
benchmark/             a separate uv project comparing rustruct against
                       struct/ctypes/dataclasses-struct/construct

Implementation status

Covered:

• all fixed-width types, raw, bytes/str/cstr, bits (MSB-first), flags (keep/strict/ignore), struct (including size windows), array (count/until_eof), switch, digest (CRC presets with Rocksoft overrides, md5/sha1/sha256, over as a name tuple or "*" with self-zeroing);
• expressions (the full operator set), lexical backward-only ref resolution, registers and span registers, limits (max, max_count, depth 64, an 8-deep Expr stack);
• pack with backpatching: derived lengths (linear inversion a*x + b, including derived bits fields), digests computed innermost-first, consistency checks across multiple consumers of one register;
• the streaming contract: parse -> Incomplete.needed with monotonic progress, distinguishing "hit a window boundary" (Invalid) from "ran out of buffer" (Incomplete);
• errors carry kind/path/offset; the path is assembled only while unwinding;
• coalescing of fixed fields (a static schema lowers to exactly one Fixed op, enforced by a test), min_size/static_size.

Known gaps and limitations:

• to_bytes/from_bytes (Program serialization) raise NotImplementedError; the cache format hasn't been designed.
• A switch discriminant is never derived from its own on= ref at pack time: the caller always supplies it explicitly, and a field cannot be both derived and a switch discriminant (SchemaError).
• byteorder accepts "network" as a struct-module-style alias for "big". "native" is forbidden, since it would make the wire format depend on the machine running the code.
• String encodings in the core: utf-8 / ascii / latin-1, errors="strict" only. Other CPython codecs are a frontend concern; the core currently raises SchemaError for anything else.
• An extra digest.algo="ip" preset (RFC 1071, the IPv4 header checksum) is needed for the bundled IPv4 reference format.
• No cargo-fuzz targets and no refcount tests yet; deterministic and randomized round-trip tests stand in for a proper proptest-based fuzz suite.
• The hot path still builds an intermediate Value tree (the core stays Python-free); raw FFI and METH_FASTCALL benchmarking haven't been done.