1. MicroPython MessagePack Introduction

MessagePack is a serialization protocol similar to JSON. Where JSON produces human readable strings, MessagePack produces binary bytes data. The protocol achieves a substantial reduction in data volume. Its combination of ease of use, extensibility, and packing performance makes it worth considering in any application where data volume is an issue.

This MicroPython implementation has usage identical to that of the ujson library:

import umsgpack
obj = [1, 2, 3.14159]
s = umsgpack.dumps(obj)  # s is a bytes object

MessagePack efficiently serialises a range of Python built-in types, a range that can readily be extended to support further Python types or instances of user defined classes. A key feature of the protocol is that, like JSON, it is self describing: the receiver gets all the information needed to decode the stream from the stream itself. This example assumes a file whose contents were encoded with MessagePack.

import umsgpack
with open('data', 'rb') as f:
    z = umsgpack.load(f)
print(z)  # z might be a wide range of Python objects

The protocol is self describing even when extended. The file umsgpack/umsgpack_ext.py extends support to complex, tuple and set built-in types. Provided that file exists on the target, the code above will work even if the data includes such objects. Extension can be provided in a way that is transparent to the application.

This document focuses on usage with MicroPython and skips most of the detail on how MessagePack works. The source document provides a lot of useful information for those wishing to understand the protocol. Also see the MessagePack spec.

1.1 Supported types

The following types are natively supported.

int In range -2**63 to 2**64 - 1.
True, False and None.
str Unicode string.
bytes Binary data.
float IEEE-754 single or double precision controlled by a dump option.
tuple By default tuples are de-serialised as lists, but this can be overridden by a load option. The extension module provides proper support wherby list and tuple objects unpack unchanged.
list Termed "array" in MessagePack documentation.
dict Termed "map" in MessagePack docs.

The list and dict types may contain any supported type including list and dict instances (ad nauseam).

The file umsgpack_ext.py (in umsgpack directory) extends this to add the following:

tuple Provides an explicit distinction between tuple and list types: if it is encoded as a tuple it will be decoded as one.
complex
set

1.2 Performance

The degree of compression compared to UJSON depends on the object being serialised. The following gives an example:

import umsgpack
import ujson
obj = [1, True, False, 0xffffffff, {u"foo": b"\x80\x01\x02", u"bar": [1,2,3, {u"a": [1,2,3,{}]}]}, -1, 2.12345]
lj = len(ujson.dumps(obj))
lm = len(umsgpack.dumps(obj, force_float_precision = "single"))
print(lj, lm, lj/lm)

Outcome: lj == 106 bytes, lm == 41 bytes corresponding to a transmission speedup factor of 2.6. The force_float_precision option ensures the same result on 32 bit and 64 bit platforms.

If large quantities of text are to be transmitted, a greater gain could be achieved by compressing the text with a gzip style compressor and serialising the resultant bytes object with MessagePack.

2. The MicroPython implementation

This implementation is based on the following well proven MessagePack repo u-msgpack-python. This runs under Python 2 and Python 3. With trivial adaptations it will run under MicroPython but at a high cost in RAM consumption. This version was adapted from that codebase and optimised to minimise RAM usage when run on microcontrollers. Consumption is about 12KiB measured on STM32. Using frozen bytecode this reduces to about 3.5KiB. This was tested with the asyntest.py demo, comparing the free RAM with that available running a similar script which exchanges uncompressed data. The difference was taken to be the library overhead of running compression and asynchronous decompression.

This version is a subset of the original. Support was removed for features thought unnecessary for microcontroller use. The principal example is that of timestamps. MicroPython does not support the datetime module. There are also issues with platforms differing in their time handling, notably the epoch. On a microcontroller it is simple to send the integer result from time.time() or even time.time_ns(). Python 2 support is removed. The code will run under Python 3.

Supported types are fully compliant with a subset of the latest MessagePack specification. In particular, it supports the new binary, UTF-8 string and application-defined ext types. As stated above, timestamps are unsupported.

The repository includes umsgpack/umsgpack_ext.py which optionally extends the library to support Python set, complex and tuple objects. The aim is to show how this can easily be extended to include further types.

This MicroPython version uses various techniques to minimise RAM use including "lazy imports": a module is only imported on first usage. For example an application that only de-serialises data using synchronous code will not import code to dump data or that to support asynchronous programming.

3. Installation

Clone the repo by moving to a directory on your PC and issuing

$ git clone https://github.com/peterhinch/micropython-msgpack

Copy the directory umsgpack and its contents to your target hardware.

The following optional files may also be copied to the target:

user_class.py A serialisable user class.
asyntest.py Demo of asynchronous use. See section 7.

The files run_test_suite and test_umsgpack.py comprise the test suite which runs on a PC. See section 9.

If RAM usage is to be minimised, the file umsgpack/umsgpack_ext.py may be deleted from the target with loss of its additional type support. Its overhead is about 2KiB measured on a Pyboard with no frozen bytecode.

4. API

In applications using ujson, MessagePack provides a drop-in replacement with the same API. The human readable (but large) strings are replaced by compact binary bytes objects.

The API supports the following methods:

dumps(obj, **options) Pack an arbitrary Python object. Returns a bytes instance.
dump(obj, fp, **options) Pack a Python object to a stream or file.
loads(s, **options) Deserialise a bytes object. Returns a Python object.
load(fp, **options) Unpack data from a stream or file.
aload(fp, **options) Asynchronous unpack of data from a StreamReader. See section 7.

The options are discussed below. Most are rather specialised. I am unsure if there is a practical use case for ext_handlers: an easier way is to use the ext_serializable decorator.

4.1 Load options

load, loads and aload support the following options as keyword args:

allow_invalid_utf8 (bool): unpack invalid strings into bytes (default False which causes an exception to be raised).
use_ordered_dict (bool): unpack dicts into OrderedDict, instead of dict. (default False).
use_tuple (bool): unpacks arrays into tuples, instead of lists (default False). The extension module (if used) makes this redundant.
ext_handlers a dictionary of Ext handlers, mapping integer Ext type to a callable that unpacks an instance of Ext into an object. See section 8.

Work is in progress to make dict instances ordered by default, so option 3 may become pointless. The umsgpack_ext module enables tuples to be encoded in a different format to lists which is more flexible than the global use_tuple arg.

4.2 Dump options

dump and dumps support the following options as keyword args:

force_float_precision (str): "single" to force packing floats as IEEE-754 single-precision floats, "double" to force packing floats as IEEE-754 double-precision floats. By default the precision of the target's firmware is detected and used.
ext_handlers (dict): dictionary of Ext handlers, mapping a custom type to a callable that packs an instance of the type into an Ext object. See section 8.

5. Extension module

The umsgpack_ext module extends umsgpack to support complex, set and tuple types, but its design facilitates adding further Python built-in types or types supported by other libraries. Support is entirely transparent to the application: the added types behave in the same way as native types.

The following examples may be pasted at the REPL:

import umsgpack
with open('data', 'wb') as f:
   umsgpack.dump(1 + 4j, f)  # mpext() handles extension type

Reading back:

import umsgpack
with open('data', 'rb') as f:
    z = umsgpack.load(f)
print(z)  # z is complex

The file umsgpack_ext.py may be found in the umsgpack directory. To extend it to support additional types, see section 11.

6. Serialisable user classes

An example of a serialisable user class may be found in user_class.py. It provides a Point3d class representing a point in 3D space stored as three float values. It may be run as follows (paste at the REPL):

import umsgpack
from user_class import Point3d
p = Point3d(1.0, 2.1, 3)
s = umsgpack.dumps(p)
print(umsgpack.loads(s))

6.1 The ext_serializable decorator

This provides a simple way of extending MessagePack to include additional types. The following is the contents of user_class.py:

import umsgpack
import struct

@umsgpack.ext_serializable(0x10)
class Point3d:
    def __init__(self, x, y, z):
        self.v = (float(x), float(y), float(z))

    def __str__(self):
        return "Point3d({} {} {})".format(*self.v)

    def packb(self):
        return struct.pack(">fff", *self.v)

    @staticmethod
    def unpackb(data):
        return Point3d(*struct.unpack(">fff", data))

A class defined with the decorator must provide the following methods:

Constructor: stores the object to be serialised.
packb This returns a bytes instance containing the serialised object.
unpackb Defined as a static method, this accepts a bytes instance of packed data and returns a new instance of the unpacked data type.

Typically this packing and unpacking is done using the struct module, but in the some simple cases it may be done by umsgpack itself. The following, taken from the extension module, illustrates support for complex:

@umsgpack.ext_serializable(0x50)
class Complex:
    def __init__(self, c):
        self.c = c

    def __str__(self):
        return "Complex({})".format(self.c)

    def packb(self):
        return struct.pack(">ff", self.c.real, self.c.imag)

    @staticmethod
    def unpackb(data):
        return complex(*struct.unpack(">ff", data))

7. Asynchronous use

7.1 Serialisation

Serialisation presents no problem in asynchronous code. The following example serialises the data using the normal synchronous dumps method then sends it asynchronously:

async def sender():
    swriter = asyncio.StreamWriter(uart, {})
    obj = [1, 2, 3.14159]
    while True:
        s = umsgpack.dumps(obj)  # Synchronous serialisation
        swriter.write(s)
        await swriter.drain()  # Asynchonous transmission
        await asyncio.sleep(5)
        obj[0] += 1

7.2 De-serialisation

This is potentially difficult. In the case of ASCII protocols like JSON and Pickle it is possible to append a b'\n' delimiter to each message, then use StreamReader.readline() to perform an asynchronous read of an entire message. This works because the messages themselves cannot contain that character. MessagePack is a binary protocol so the data may include all possible byte values. Consequently a unique delimiter is unavailable.

MessagePack messages are binary sequences whose length is unknown to the receiver. Further, in many case a substantial amount of the message must be read before the total length can be deduced. The solution adopted is to add an aload() method that accepts data from a StreamReader, decoding it as it arrives. The following is an example of an asynchronous reader:

async def receiver():
    sreader = asyncio.StreamReader(uart)
    while True:
        res = await umsgpack.aload(sreader)
        print('Recieved', res)

The demo asyntest.py runs on a Pyboard with pins X1 and X2 linked. The code includes notes regarding RAM overhead.

8. Ext Handlers

This is an alternative to the ext_serializable decorator and provides another option for extending MessagePack. In my view it is rather clunky and I struggle to envisage a use case. It is included for completeness.

The packing functions accept an optional ext_handlers dictionary that maps custom types to callables that pack the type into an Ext object. The callable should accept the custom type object as an argument and return a packed umsgpack.Ext object.

Example for packing set and complex types into Ext objects with type codes 0x20 and 0x30:

umsgpack.dumps([1, True, {"foo", 2}, complex(3, 4)],
    ext_handlers = {
       set: lambda obj: umsgpack.Ext(0x20, umsgpack.dumps(list(obj))),
       complex: lambda obj: umsgpack.Ext(0x30, struct.pack("ff", obj.real, obj.imag))
       })

Similarly, the unpacking functions accept an optional ext_handlers dictionary that maps Ext type codes to callables that unpack the Ext into a custom object. The callable should accept a umsgpack.Ext object as an argument and return an unpacked custom type object.

Example for unpacking Ext objects with type codes 0x20, and 0x30 into set and complex objects:

umsgpack.loads(s,
    ext_handlers = {
      0x20: lambda ext: set(umsgpack.loads(ext.data)),
      0x30: lambda ext: complex(*struct.unpack("ff", ext.data)),
    })

Example for packing and unpacking a custom class:

class Point(object):
    def __init__(self, x, y, z):
        self.x = x
        self.y = y
        self.z = z

    def __str__(self):
        return "Point({}, {}, {})".format(self.x, self.y, self.z)

    def pack(self):
        return struct.pack(">iii", self.x, self.y, self.z)

    @staticmethod
    def unpack(data):
        return Point(*struct.unpack(">iii", data))

# Pack
obj = Point(1,2,3)
data = umsgpack.dumps(obj, ext_handlers = {Point: lambda obj: umsgpack.Ext(0x10, obj.pack())})

# Unpack
obj = umsgpack.loads(data, ext_handlers = {0x10: lambda ext: Point.unpack(ext.data)})

9. Exceptions

These are defined in umsgpack/__init__.py.

The dump and dumps methods can throw the following:

# Packing error
class UnsupportedTypeException(PackException):
    "Object type not supported for packing."

The load and loads methods can throw the following. In practice these are only likely to occur if data has been corrupted, for example if transmitted via an unreliable medium:

class InsufficientDataException(UnpackException):
    "Insufficient data to unpack the serialized object."

class InvalidStringException(UnpackException):
    "Invalid UTF-8 string encountered during unpacking."

class ReservedCodeException(UnpackException):
    "Reserved code encountered during unpacking."

class UnhashableKeyException(UnpackException):
    """
    Unhashable key encountered during map unpacking.
    The serialized map cannot be deserialized into a Python dictionary.
    """

class DuplicateKeyException(UnpackException):
    "Duplicate key encountered during map unpacking."

10. Test suite

This is mainly of interest to those wanting to modify the code.

The repo includes the test suite test_umsgpack.py which must be run under Python 3 in a directory containing the umsgpack tree. It will not run under MicroPython because it tests large data types: the test suite causes memory errors when compiled under even the Unix build of MicroPython. The file umsgpack_ext.py should not be present: this is because the test suite assumes that complex and set are not supported. The script run_test_suite renames umsgpack_ext.py, runs the tests and restores the file.

11. Changes for MicroPython

Code in this repo is based on this implementation, whose code is of high quality. It required minimal changes to run under MicroPython.

Hosted on a microcontroller its RAM consumption was high; most changes were to reduce this. Further, the nature of the protocol results in an issue when using asynchronous code to de-serialise data which arrives slowly or sporadically. This is handled by adding an asynchronous de-serialiser.

These can be summarised as follows:
Python2 code removed.
Compatibility mode removed.
Timestamps removed.
Converted to Python package with lazy import to save RAM.
Provide uasyncio StreamReader support.
Exported functions now match ujson: dump, dumps, load, loads (only).
Many functions refactored to save bytes, e.g. replacing the function dispatch table with code.
Further refactoring to reduce allocations.
InvalidString class removed because it is a subclass of a native type.
Method of detecting platform's float size changed (MicroPython does not support the original method).
Version reset to (0.1.0).

12. Notes on the extension module

These notes are for those wishing to understand how this works, perhaps to add support for further types.

The mp_dump.py attempts to load a function mpext from the module. If this fails (becuase the module is missing) it creates a dummy function. When the dump method runs, it executes mpext passing the object to be encoded. If the type of the object matches one support by the extension, it returns an instance of a serialisable class created with the passed object. If the type does not match, the passed object is returned for dump to inspect.

Supporting additional types therefore comprises the following:

Create an ext_serializable class for the new type as per section 6.1.
Change the function mpext to check for the new type and, if found, return an instance of the above class.

Acknowledgements

This project was inspired by this forum thread where user WZab performed an initial port of the source library. See also this GitHub issue.

Summary of references

MessagePack main site:
MessagePack
MessagePack spec:
the MessagePack spec.
Code on which this repo is based:
u-msgpack-python.

License

micropython-msgpack is MIT licensed. See the included LICENSE file for more details.

facchinm/micropython-msgpack