2 unstable releases
0.2.0 | Jul 11, 2021 |
---|---|
0.1.0 | Jul 9, 2021 |
#1140 in Procedural macros
9KB
112 lines
pyo3-prost
pyo3-prost exposes Rust protobuf structs to Python.
Docs []
# A simple decoding benchmark
Python: 14.008996680990094
Rust : 0.5708326659951126
# A simple encoding benchmark
Python: 8.68384731200058
Rust : 0.3029898940003477
===========
PythonTweet
===========
b'\n\x11Hi this is a text\x10\xfb\x99K"\x1b\n\x03Who\x12\x14https://example.com/*\x06@trump*\x06@obama'
===========
RustTweet
===========
b'\n\x11Hi this is a text\x10\xfb\x99K"\x1b\n\x03Who\x12\x14https://example.com/*\x06@trump*\x06@obama'
All you need is adding a single line
.type_attribute(".", "#[::pyo3_prost::pyclass_for_prost_struct]")
to your prost or tonic build.
The intended use-case is when your Python project is locked down to Protobuf and you want to migrate some part to native modules in Rust.
Python protobuf objects cannot be easily handled in the Rust side.
For conversion, the Python protobuf object should be serialized and deserialized to a Rust protobuf object.
With pyo3-prost
you can decode the bytes to Rust structs in Python from the beginning and pass it to the native module.
How it works
The derive macro pyclass_for_prost_struct
will add
#[pyclass]
to prost-generated structs#[pyo3(get, set)]
for each field (except oneof fields).#[pymethods]
withencode() -> &PyBytes
,decode(&PyBytes) -> Self
,decode_merge(&mut self, &PyBytes)
.#[pyproto]
with__str__
and__repr__
for easy inspection.
- Getter/setter works with nested structs.
encode_slow
is name this way because it's currently implemented by serializing to Rust buffer and copy the bytes to Python memory.
Example
- Go to
examples/proto-gen
andcargo run
. This will createexamples/rupy_proto/src/app.rs
. - Go to
examples/rupy_proto
andcargo build --release
. - Move
examples/rupy_proto/target/release/librupy_proto.so
toexamples/rupy_proto/rupy_proto.so
. - Run the bench
cd examples/rupy_proto && PYTHONPATH=. python bench.py
.
Limitations (for now)
Accessing numeric fields is fast. However each access to a repeated, map, or message field may return a cloned Python object.
For example obj.list_field.clear()
will only clear the returned copy, leaving the actual value untouched.
Getters/setters are not provided for oneof
fields.
Later I will add a custom getter/setter for these cases.
Dependencies
~1.5MB
~36K SLoC