marshmallow-core 0.1.7

Rust acceleration core for marshmallow serialization, installed as a separate, opt-in package.

marshmallow_core

A Rust acceleration core for marshmallow, shipped as a separate, opt-in package. Install it next to stock marshmallow and activate it explicitly — it replaces marshmallow's per-object _serialize / _deserialize loops with a PyO3 extension while producing identical results.

pip install marshmallow marshmallow_core

import marshmallow as ma
import marshmallow_core

marshmallow_core.install()      # patch marshmallow.Schema in this process

class Person(ma.Schema):
    name = ma.fields.String()
    age = ma.fields.Integer()

Person().load({"name": "ann", "age": "30"})   # accelerated
Person().dump({"name": "ann", "age": 30})      # accelerated

marshmallow_core.uninstall()    # restore the stock pure-Python methods

How it works

• install() monkey-patches Schema._serialize and Schema._do_load. Each bound schema is compiled once (cached on the instance) into a recursive payload describing every field as either native (run entirely in Rust) or a callback (defers to the Python Field method). Anything not modelled natively stays a callback, so output is behaviour-identical.
• Both cores handle the happy path and raise an internal AccelFallback on any error/edge case, so marshmallow re-runs the unchanged pure-Python path and every error message and value matches exactly. (Dump has no side effects — it builds a fresh output — so it can discard a partial result and re-run safely, just like load.)
• dumps is fused: it writes JSON bytes directly in Rust, skipping the intermediate Python dict and the json.dumps pass, byte-for-byte identical to json.dumps(schema.dump(obj)). It activates for hook-free schemas using the stdlib json render module with no extra json kwargs, and falls back to dump + json.dumps for anything it can't reproduce exactly. (loads is already accelerated through the patched load path; a Rust JSON parser was prototyped but did not beat CPython's C json.loads, so it was not shipped.)
• Acceleration is strictly a speedup. Set MARSHMALLOW_NO_ACCEL=1 (or hit a protocol-version mismatch between the Python and Rust halves) and the core becomes a no-op even after install().

Scope / limitations

install() accelerates dump for all compilable schemas, and load for most schemas — including those with pre_load / post_load / validates / validates_schema hooks: the core runs the per-field deserialize step while those hooks run in Python around it (mirroring marshmallow's own _do_load split). Recognized field validators (Range / Length / OneOf / Equal / NoneOf / ContainsOnly) run natively; any other validator, or field-level pre_load / post_load, keeps that field on the callback path. unknown=INCLUDE, collection/dotted partial, and dotted attribute writes are all accelerated. Natively modelled fields include the scalars plus Decimal, Dict (incl. typed keys/values), Tuple, Pluck, Constant, TimeDelta, Boolean, Integer(strict=True), and NaiveDateTime / AwareDateTime. The dump core has an AccelFallback (it discards a partial result and re-runs pure Python on any shape it can't reproduce), so it accelerates the composite fields too. Custom dict_class / get_attribute, self-referential schemas, custom strptime temporal formats, and callable defaults always fall back to pure Python.

Where the speedup is limited

Some shapes are inherently bounded — the work the core can't move into Rust dominates the call. These are correct, just not where the gains are:

• Hook-bearing loads are the weakest case (~2x, vs ~7x without hooks). When a schema has pre_load / post_load / validates / validates_schema, the core runs the per-field deserialize but marshmallow's Python hook-dispatch (_invoke_load_processors, _invoke_field_validators) runs around it. On a small schema the core step is ~8% of the load; the remaining ~90% is that Python machinery, which wraps user callbacks and cannot be moved into Rust.
• Small / flat schemas are capped by fixed per-call overhead (~20–30%). The dump / load entry prologue (argument normalization, the per-instance serializer-cache lookup, the partial/unknown checks) is constant per call, so it dominates exactly when the payload is tiny. It amortizes to near-zero as the payload grows — speedup on a list of records is flat regardless of length.
• loads gains less than load. JSON parsing still goes through CPython's C json.loads; a fused Rust parser was prototyped but did not beat it, so only the subsequent per-field step is accelerated.

For collections of records (the common hot path) the fixed overhead vanishes and the speedup is steady (~7–8x). Run performance/analyze_paths.py to see whether a given schema even reaches the core, and performance/benchmark.py to measure it.

Development

Requires cargo (rustup) and maturin.

# build + install the extension into the current venv
uvx maturin develop --release

# run the tests (needs marshmallow + pytest installed)
pytest

# force the pure-Python path
MARSHMALLOW_NO_ACCEL=1 pytest

tests/test_equivalence.py asserts that dump/load produce identical output and errors with the core active vs. forced onto pure Python, across scalars, nested/list/enum/temporal/UUID fields, partial=True, and error inputs.

Benchmarking

The performance/ directory (not shipped in wheels) measures the core against stock marshmallow through the public install() / uninstall() API. Run it from the repo root with the compiled extension importable (uvx maturin develop --release first, or point PYTHONPATH at the repo while the wheel is installed):

# stock-vs-core table for dump / load / dumps / loads on four schema shapes
python -m performance.benchmark                       # all cases
python -m performance.benchmark --number 20000 --only flat,list

# coverage probe: per-field native vs callback for each schema shape
python -m performance.analyze_paths

benchmark.py reports per-call microseconds for stock and core plus the speedup ratio, across flat-scalar, nested, list-heavy, and validator-heavy schemas. analyze_paths.py inspects the compiled payload and shows which fields run native in Rust vs. fall back to a Python callback — it tells you exactly where a real schema still defers to pure Python.

Releasing

CI (.github/workflows/ci.yml) builds the wheel and runs the suite against stock marshmallow on Python 3.10–3.13, both with the core active and with MARSHMALLOW_NO_ACCEL=1. Publishing (.github/workflows/release.yml) builds abi3 wheels + sdist for Linux/macOS/Windows on a v* tag and uploads them to PyPI via trusted publishing. Before the first release, configure the PyPI trusted publisher for this repo and create a pypi GitHub Environment, then push a tag (e.g. git tag v0.1.0 && git push --tags).

License

MIT