tmsgpack 0.1.3

Typed MessagePack serializer (similar but incompatible with msgpack)

tmsgpack: Typed MessagePack-inspired pack/unpack component

See also: FORMAT.md, DEVELOP.md tmsgpack-js

The tmsgpack format expresses typed objects: maps and arrays (or: dicts and tuples/lists) with an object_type property.

Unlike msgpack and pickle, this is not a batteries-included end-to-end serialization solution. It is a composable component that helps you to build end-to-end communication solutions.

Your system solution design will make decisions on:

• What objects are serializable and what objects are not.
• What code to use (and, maybe, dynamically load) to instantiate serialized objects.
• How to represent objects that are unpacked but not supposed to 'live' in this process.
• How to share dynamic data between different packs/unpacks.
• How to asynchronously load and integrate shared data from different sources.
• How to map typed object meaning between different programming languages.
• Whether and how to convert persisted "old" data to current, new semantics (schemas).
• How much to attach explicit meaning and predictable schemas to your object types.
• Whether or not to use the 'expression execution' capabilities of tmsgpack.
• etc.

This python package makes a minimal (backwards-incompatible) modification to the msgpack format to make all this elegantly possible. This package is based on msgpack v1.0.5.

Installation

pip install tmsgpack

Usage

Packing and unpacking data is controlled by pack_ctrl and unpack_ctrl objects (see below for details how to create them):

from tmsgpack import packb, unpackb
packed = packb(data, pack_ctrl=pack_ctrl)
unpacked = unpackb(packed, unpack_ctrl=unpack_ctrl)

Streaming unpacking

For multiple uses, you can use packer and unpacker objects:

from tmsgpack import Packer
packer = Packer(pack_ctrl=pack_ctrl)

packed = packer.pack(data) # Send these packages via a socket...

---
from tmsgpack import Packer, Unpacker

unpacker = Unpacker(unpack_ctrl=unpack_ctrl)
while buf := sock.recv(1024**2):
    unpacker.feed(buf)
    for o in unpacker:
        process(o)

Minimal pack_ctrl and unpack_ctrl objects

Minimal controllers allow only JSON-like objects and raise errors when you ask for more (below, we show examples for more useful controllers):

from tmsgpack import PackConfig, UnpackConfig
from dataclasses import dataclass

@dataclass
class MinimalPackCtrl:
    def from_obj(self, obj):
        raise TypeError(f'Cannot serialize {type(obj)} object.')
    options: PackConfig

@dataclass
class MinimalUnpackCtrl:
    def from_dict(self, ctype, dct):
        raise ValueError(f'Unpack type not supported: {ctype} data: {dct}')
    def from_tuple(self, ctype, lst):
        raise ValueError(f'Unpack type not supported: {ctype} data: {lst}')
    options: UnpackConfig

def pctrl(**kwargs): return MinimalPackCtrl(options=PackConfig(**kwargs))
def uctrl(**kwargs): return MinimalUnpackCtrl(options=UnpackConfig(**kwargs))

minimal_pack_ctrl = pctrl()
minimal_unpack_ctrl = uctrl()

The API and configuration

As you see, the pack_ctrl object provides a method from_obj. The unpack_ctrl object provides the methods from_dict and from_tuple:

as_dict, data_type, data = pack_ctrl.from(obj)

# When `as_dict` is true, then `data` should be a dictionary.
# When `as_dict` is false, then `data` should be a tuple or a list.

unpacked = unpack_ctrl.from_dict(data_type, data) # used when as_dict is true.
unpacked = unpack_ctrl.from_tuple(data_type, data) # used when as_dict is false.

PackConfig configuration objects for pack_ctrl

PackConfig objects provide the following options:

from tmsgpack import PackConfig

config = PackConfig(
    p_shortcuts=True, sort_keys=False, p_str_keys=False, use_single_float=False,
)
"""
Config object for pack_ctrl.options

    :param bool p_shortcuts:
        If true, pack dicts, tuples, lists without calling from_obj(...).
        Object types: {...} => None; (...) => None; [...] => False
        (default: True)

    :param bool sort_keys:
        Sort output dictionaries by key. (default: False)

    :param bool p_str_keys:
        If true, accept only strings as dict keys. (default: False)

    :param bool use_single_float:
        Use single precision float type for float. (default: False)
"""

UnpackConfig configuration objects for unpack_ctrl

UnpackConfig objects provide the following options:

from tmsgpack import UnpackConfig

config = UnpackConfig(
    u_shortcuts=True, u_str_keys=False,
    read_size=16*1024, max_buffer_size=0,
    max_str_len=-1, max_bin_len=-1, max_list_len=-1, max_dict_len=-1,
)
"""
Config object for unpack_ctrl.options

    :param bool u_shortcuts:
        If true, unpack dicts, tuples, lists without calling from_dict/from_tuple.
        Object types: None => {...}/(...), False => [...]
        (default: True)

    :param bool u_str_keys:
        If true, accept only strings as dict keys. (default: False)

    :param int read_size:
        Used as `file_like.read(read_size)`. (default: `min(16*1024, max_buffer_size)`)

    :param int max_buffer_size:
        (default: 100*1024*1024 (100MiB))
        Limits size of data waiting unpacked.  0 means 2**32-1.
        Raises `BufferFull` exception when it is insufficient.
        You should set this parameter when unpacking data from untrusted source.

    :param int max_str_len:
        Limits max length of str. (default: max_buffer_size)

    :param int max_bin_len:
        Limits max length of bin. (default: max_buffer_size)

    :param int max_list_len:
        Limits max length of list.
        (default: max_buffer_size)

    :param int max_dict_len:
        Limits max length of dict.
        (default: max_buffer_size//2)
"""

Packing and Unpacking dataclass objects

Here are the parts of one unit test that shows end-to-end packing and unpacking of dataclass objects:

For the setup, we import tools and define the controllers:

from tmsgpack import packb, unpackb, PackConfig, UnpackConfig
from dataclasses import dataclass, is_dataclass, fields
from typing import Dict

@dataclass
class TypedPackCtrl:
    def from_obj(self, obj):
        if type(obj) is tuple: return False, 'tuple', obj  # Special case for tuples.
        if not is_dataclass(obj): raise TypeError(f'Cannot serialize {type(obj)} object.')
        as_dict = not getattr(obj, 'as_list', False)
        object_type = obj.__class__.__name__
        if as_dict:
            data = {
                field.name: getattr(obj, field.name)
                for field in fields(obj)
            }
        else:
            data = [
                getattr(obj, field.name)
                for field in fields(obj)
            ]
        return as_dict, object_type, data
    options: PackConfig

    def pack(self, data):
        return packb(data, pack_ctrl=self)

@dataclass
class TypedUnpackCtrl:
    constructors: Dict[str, callable]
    def from_dict(self, ctype, data): return self.constructors[ctype](**data)
    def from_tuple(self, ctype, data): return self.constructors[ctype]( *data)
    options: UnpackConfig

    def unpack(self, packed):
        return unpackb(packed, unpack_ctrl=self)

def pc(**kwargs): return TypedPackCtrl(options=PackConfig(**kwargs))
def uc(fns, **kwargs):
    return TypedUnpackCtrl(
        constructors={fn.__name__:fn for fn in fns},
        options=UnpackConfig(**kwargs),
    )

Notes:

• For conveninence, we added methods pack_ctrl.pack(data) and unpack_ctrl.unpack(packed).
• The method pack_ctrl.from_obj decides whether to represent the dataclass object as a key-value dict or as a more compact list of values.
• It extracts the properties of the obj and sets the values as_dict, object_type and data appropriately.
• In this implementation, object types are the unqualified class names. There is a possibility that one class from one package can have the same name as a different class from a different package.
• Fully resolving naming spaces is a deep design problem. You need to decide what you mean by 'meaning'. Here, we exploit this overloadability...
• The unpack_ctrl object is created with a list of available constructor functions.

Now, we can define the data classes to be packed and unpacked:

@dataclass
class Foo:
    x: int = 1
    y: str = 'Y'

@dataclass
class Bar:
    x: int = 1
    y: str = 'Y'
    as_list = True

Notes:

• Class Bar has a class property as_list=True. It will be packed compactly as a parameter value list.

Here is a simple test runner to be used for several tests:

def run(input, expected=None):
    if expected is None: expected = input

    constructors = [Foo, Bar]

    pack_ctrl    = pc()
    unpack_ctrl  = uc(constructors)

    packed = pack_ctrl.pack(input)
    output = unpack_ctrl.unpack(packed)

    assert output == expected

The Foo() and Bar() objects are packed and unpacked correctly:

def test_typed_foobar():
    run(Foo())   # Encoded as a typed dict
    run(Bar())   # Encoded as a typed list
    run((1,2,3)) # Tuples are encoded as a typed list with object_type=None.
    run([1,2,3]) # Tuples are encoded as a typed list with object_type=False.
    run({'hello':'world'}) # Dicts are encoded as a typed dict with object_type=None.