Messages

The fundamental building block in protocol buffers are messages. Messages are essentially permissive, strongly-typed structs (dictionaries), which have zero or more fields that may themselves contain primitives or other messages.

syntax = "proto3";

message Song {
  Composer composer = 1;
  string title = 2;
  string lyrics = 3;
  int32 year = 4;
}

message Composer {
  string given_name = 1;
  string family_name = 2;
}

The most common use case for protocol buffers is to write a .proto file, and then use the protocol buffer compiler to generate code for it.

Declaring messages

However, it is possible to declare messages directly. This is the equivalent message declaration in Python, using this library:

import proto

class Composer(proto.Message):
    given_name = proto.Field(proto.STRING, number=1)
    family_name = proto.Field(proto.STRING, number=2)

class Song(proto.Message):
    composer = proto.Field(Composer, number=1)
    title = proto.Field(proto.STRING, number=2)
    lyrics = proto.Field(proto.STRING, number=3)
    year = proto.Field(proto.INT32, number=4)

A few things to note:

This library only handles proto3.
The number is really a field ID. It is not a value of any kind.
All fields are optional (as is always the case in proto3). The only general way to determine whether a field was explicitly set to its falsy value or not set all is to mark it optional.
Because all fields are optional, it is the responsibility of application logic to determine whether a necessary field has been set.
You can optionally define a __protobuf__ attribute in your module which will be used to differentiate messages which have the same name but exist in different modules.

# file a.py
import proto

__protobuf__ = proto.module(package="a")

class A(proto.Message):
    name = proto.Field(proto.STRING, number=1)

# file b.py
import proto

__protobuf__ = proto.module(package="b")

class A(proto.Message):
    name = proto.Field(proto.STRING, number=1)

# file main.py
import a
import b

_a = a.A(name="Hello, A!")
_b = b.A(name="Hello, B!")

Messages are fundamentally made up of Fields. Most messages are nothing more than a name and their set of fields.

Usage

Instantiate messages using either keyword arguments or a dict (and mix and matching is acceptable):

>>> song = Song(
...     composer={'given_name': 'Johann', 'family_name': 'Pachelbel'},
...     title='Canon in D',
...     year=1680,
... )
>>> song.composer.family_name
'Pachelbel'
>>> song.title
'Canon in D'

Assigning to Fields

One of the goals of proto-plus is to make protobufs feel as much like regular python objects as possible. It is possible to update a message’s field by assigning to it, just as if it were a regular python object.

song = Song()
song.composer = Composer(given_name="Johann", family_name="Bach")

# Can also assign from a dictionary as a convenience.
song.composer = {"given_name": "Claude", "family_name": "Debussy"}

# Repeated fields can also be assigned
class Album(proto.Message):
    songs = proto.RepeatedField(Song, number=1)

a = Album()
songs = [Song(title="Canon in D"), Song(title="Little Fugue")]
a.songs = songs

Note

Assigning to a proto-plus message field works by making copies, not by updating references. This is necessary because of memory layout requirements of protocol buffers. These memory constraints are maintained by the protocol buffers runtime. This behavior can be surprising under certain circumstances, e.g. trying to save an alias to a nested field.

proto.Message defines a helper message, copy_from() to help make the distinction clear when reading code. The semantics of copy_from() are identical to the field assignment behavior described above.

composer = Composer(given_name="Johann", family_name="Bach")
song = Song(title="Tocatta and Fugue in D Minor", composer=composer)
composer.given_name = "Wilhelm"

# 'composer' is NOT a reference to song.composer
assert song.composer.given_name == "Johann"

# We CAN update the song's composer by assignment.
song.composer = composer
composer.given_name = "Carl"

# 'composer' is STILL not a reference to song.composer.
assert song.composer.given_name == "Wilhelm"

# It does work in reverse, though,
# if we want a reference we can access then update.
composer = song.composer
composer.given_name = "Gottfried"

assert song.composer.given_name == "Gottfried"

# We can use 'copy_from' if we're concerned that the code
# implies that assignment involves references.
composer = Composer(given_name="Elisabeth", family_name="Bach")
# We could also do Message.copy_from(song.composer, composer) instead.
Composer.copy_from(song.composer, composer)

assert song.composer.given_name == "Elisabeth"

Enums

Enums are also supported:

import proto

class Genre(proto.Enum):
    GENRE_UNSPECIFIED = 0
    CLASSICAL = 1
    JAZZ = 2
    ROCK = 3

class Composer(proto.Message):
    given_name = proto.Field(proto.STRING, number=1)
    family_name = proto.Field(proto.STRING, number=2)

class Song(proto.Message):
    composer = proto.Field(Composer, number=1)
    title = proto.Field(proto.STRING, number=2)
    lyrics = proto.Field(proto.STRING, number=3)
    year = proto.Field(proto.INT32, number=4)
    genre = proto.Field(Genre, number=5)

All enums must begin with a 0 value, which is always the default in proto3 (and, as above, indistuiguishable from unset).

Enums utilize Python enum.IntEnum under the hood:

>>> song = Song(
...     composer={'given_name': 'Johann', 'family_name': 'Pachelbel'},
...     title='Canon in D',
...     year=1680,
...     genre=Genre.CLASSICAL,
... )
>>> song.genre
<Genre.CLASSICAL: 1>
>>> song.genre.name
'CLASSICAL'
>>> song.genre.value
1

Additionally, it is possible to provide strings or plain integers:

>>> song.genre = 2
>>> song.genre
<Genre.JAZZ: 2>
>>> song.genre = 'CLASSICAL'
<Genre.CLASSICAL: 1>

Serialization

Serialization and deserialization is available through the serialize() and deserialize() class methods.

The serialize() method is available on the message classes only, and accepts an instance:

serialized_song = Song.serialize(song)

The deserialize() method accepts a bytes, and returns an instance of the message:

song = Song.deserialize(serialized_song)

JSON serialization and deserialization are also available from message classes via the to_json() and from_json() methods.

json = Song.to_json(song)

new_song = Song.from_json(json)

Similarly, messages can be converted into dictionaries via the to_dict() helper method. There is no from_dict() method because the Message constructor already allows construction from mapping types.

song_dict = Song.to_dict(song)

new_song = Song(song_dict)

Note

Although Python’s pickling protocol has known issues when used with untrusted collaborators, some frameworks do use it for communication between trusted hosts. To support such frameworks, protobuf messages can be pickled and unpickled, although the preferred mechanism for serializing proto messages is serialize().

Multiprocessing example:

import proto
from multiprocessing import Pool

class Composer(proto.Message):
   name = proto.Field(proto.STRING, number=1)
   genre = proto.Field(proto.STRING, number=2)

composers = [Composer(name=n) for n in ["Bach", "Mozart", "Brahms", "Strauss"]]

with multiprocessing.Pool(2) as p:
   def add_genre(comp_bytes):
       composer = Composer.deserialize(comp_bytes)
       composer.genre = "classical"
       return Composer.serialize(composer)

updated_composers = [
   Composer.deserialize(comp_bytes)
   for comp_bytes in p.map(add_genre, (Composer.serialize(comp) for comp in composers))
]