micropython-samples/SERIALISATION.md

Serialisation

These notes are a discussion of the serialisation libraries available to MicroPython plus a tutorial on the use of a library supporting Google Protocol Buffers (here abbreviated to protobuf). The aim is not to replace official documentation but to illustrate the relative merits and drawbacks of the various approaches.

Main readme

1. The problem

The need for serialisation arises whenever data must be stored on disk or communicated over an interface such as a socket, a UART or such interfaces as I2C or SPI. All these require the data to be presented as linear sequences of bytes. The problem is how to convert an arbitrary Python object to such a sequence, and how subsequently to restore the object.

There are numerous standards for achieving this, five of which are readily available to MicroPython. Each has its own advantages and drawbacks. In two cases the encoded strings aim to be human readable and comprise ASCII characters. In the others they comprise binary bytes objects where bytes can take all possible values. The following are the formats with MicroPython support:

1. ujson (ASCII, official)
2. pickle (ASCII, official)
3. ustruct (binary, official)
4. MessagePack binary, unofficial
5. protobuf binary, unofficial

The ujson and pickle formats produce human-readable byte sequences. These aid debugging. The use of ASCII data means that a delimiter can be used to identify the end of a message. This is because it is possible to guarantee that the delimiter will never occur within a message. A delimiter cannot be used with binary formats because a message byte can take all possible values including that of the delimiter. The drawback of ASCII formats is inefficiency: the byte sequences are relatively long.

Numbers 1, 2 and 4 are self-describing: the format includes a definition of its structure. This means that the decoding process can re-create the object in the absence of information on its structure, which may therefore change at runtime. Self describing formats inevitably are variable length. This is no problem where data is being saved to file, but if it is being communicated across a link the receiving process needs a means to determine when a complete message has been received. In the case of ASCII formats a delimiter may be used but in the case of MessagePack this presents something of a challenge.

The ustruct format is binary: the byte sequence comprises binary data which is neither human readable nor self-describing. The problem of message framing is solved by hard coding a fixed message structure and length which is known to transmitter and receiver. In simple cases of fixed format data, ustruct provides a simple, efficient solution.

In protobuf and MessagePack messages are variable length; both can handle data whose length varies at runtime. MessagePack also allows the message structure to change at runtime. It is also extensible to enable the efficient coding of additional Python types or instances of user defined classes.

The protobuf standard requires transmitter and receiver to share a schema which defines the message structure. Message length may change at runtime, but structure may not.

There has been some discussion of supporting CBOR. There is a MicroPython library here. This is a binary format with a focus on minimising message length. I have not yet had time to study this.

1.1 Transmission over unreliable links

Consider a system where a transmitter periodically sends messages to a receiver over a communication link. An aspect of the message framing problem arises if that link is unreliable, meaning that bytes may be lost or corrupted in transit. In the case of ASCII formats with a delimiter the receiver, once it has detected the problem, can discard characters until the delimiter is received and then wait for a complete message.

In the case of binary formats it is generally impossible to re-synchronise to a continuous stream of data. In the case of regular bursts of data a timeout can be used. Otherwise "out of band" signalling is required where the receiver signals the transmitter to request retransmission.

1.2 Concurrency

In asyncio systems the transmitter presents no problem. A message is created using synchronous code, then transmitted using asynchronous code typically with a StreamWriter. In the case of ASCII protocols a delimiter - usually b"\n" is appended.

In the case of ASCII protocols the receiver can use StreamReader.readline() to await a complete message.

ustruct also presents a simple case in that the number of expected bytes is known to the receiver which simply awaits that number.

The variable length binary protocols present a difficulty in that the message length is unknown in advance. A solution is available for MessagePack.

2. ujson and pickle

These are very similar. ujson is documented here. pickle has identical methods so this doc may be used for both.

The advantage of ujson is that JSON strings can be accepted by CPython and by other languages. The drawback is that only a subset of Python object types can be converted to legal JSON strings; this is a limitation of the JSON specification.

The advantage of pickle is that it will accept any Python object except for instances of user defined classes. The extremely simple source may be found in the official library. The strings produced are incompatible with CPython's pickle, but can be decoded in CPython by using the MicroPython decoder. There is a bug in the MicroPython implementation when running under MicroPython. A workround consists of never encoding short strings which change repeatedly.

2.1 Usage examples

These may be copy-pasted to the MicroPython REPL.
Pickle:

import pickle
data = {1:'test', 2:1.414, 3: [11, 12, 13]}
s = pickle.dumps(data)
print('Human readable data:', s)
v = pickle.loads(s)
print('Decoded data (partial):', v[3])

JSON. Note that dictionary keys must be strings:

import ujson
data = {'1':'test', '2':1.414, '3': [11, 12, 13]}
s = ujson.dumps(data)
print('Human readable data:', s)
v = ujson.loads(s)
print('Decoded data (partial):', v['3'])

2.2 Strings are variable length

In real applications the data, and hence the string length, vary at runtime. The receiving process needs to know when a complete string has been received or read from a file. In practice ujson and pickle do not include newline characters in encoded strings. If the data being encoded includes a newline, it is escaped in the string:

import ujson
data = {'1':b'test\nmore', '2':1.414, '3': [11, 12, 13]}
s = ujson.dumps(data)
print('Human readable data:', s)
v = ujson.loads(s)
print('Decoded data (partial):', v['1'])

If this is pasted at the REPL you will observe that the human readable data does not have a line break (while the decoded data does). Output:

Human readable data: {"2": 1.414, "3": [11, 12, 13], "1": "test\nmore"}
Decoded data (partial): test
more

Consequently encoded strings may be separated with '\n' before saving and reading may be done using readline methods as in this code fragment where u is a UART instance:

s = ujson.dumps(data)
# pickle produces a bytes object whereas ujson produces a string
# In the case of ujson it is probably wise to convert to bytes
u.write(s.encode())
# Pickle:
# u.write(s)
u.write(b'\n')

# receiver
s = u.readline()
v = ujson.loads(s)  # ujson can cope with bytes object
# ujson and pickle can cope with trailing newline.

3. ustruct

This is documented here. The binary format is efficient, but the format of a sequence cannot change at runtime and must be "known" to the decoding process. Records are of fixed length. If data is to be stored in a binary random access file, the fixed record size means that the offset of a given record may readily be calculated.

Write a 100 record file. Each record comprises three 32-bit integers:

import ustruct
fmt = 'iii'  # Record format: 3 signed ints
rlen = ustruct.calcsize(fmt)  # Record length
buf = bytearray(rlen)
with open('myfile', 'wb') as f:
    for x in range(100):
        y = x * x
        z = x * 10
        ustruct.pack_into(fmt, buf, 0, x, y, z)
        f.write(buf)

Read record no. 10 from that file:

import ustruct
fmt = 'iii'
rlen = ustruct.calcsize(fmt)  # Record length
buf = bytearray(rlen)
rnum = 10  # Record no.
with open('myfile', 'rb') as f:
    f.seek(rnum * rlen)
    f.readinto(buf)
    result = ustruct.unpack_from(fmt, buf)
print(result)

Owing to the fixed record length, integers must be constrained to fit the length declared in the format string.

Binary formats cannot use delimiters as any delimiter character may be present in the data - however the fixed length of ustruct records means that this is not a problem.

For performance oriented applications, ustruct is the only serialisation approach which can be used in a non-allocating fashion, by using pre-allocated buffers as in the above example.

3.1 Strings

In ustruct the s data type is normally prefixed by a length (defaulting to 1). This ensures that records are of fixed size, but is potentially inefficient as shorter strings will still occupy the same amount of space. Longer strings will silently be truncated. Short strings are packed with zeros.

import ustruct
fmt = 'ii30s'
rlen = ustruct.calcsize(fmt)  # Record length
buf = bytearray(rlen)
ustruct.pack_into(fmt, buf, 0, 11, 22, 'the quick brown fox')
ustruct.unpack_from(fmt, buf)
ustruct.pack_into(fmt, buf, 0, 11, 22, 'rats')
ustruct.unpack_from(fmt, buf)  # Packed with zeros
ustruct.pack_into(fmt, buf, 0, 11, 22, 'the quick brown fox jumps over the lazy dog')
ustruct.unpack_from(fmt, buf)  # Truncation

Output:

(11, 22, b'the quick brown fox\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00')
(11, 22, b'rats\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00')
(11, 22, b'the quick brown fox jumps over')

4. MessagePack

Of the binary formats this is the easiest to use and can be a "drop in" replacement for ujson as it supports the same four methods dump, dumps, load and loads. An application might initially be developed with ujson, the protocol being changed to MessagePack later. Creation of a MessagePack string can be done with:

import umsgpack
obj = [1.23, 2.56, 89000]
msg = umsgpack.dumps(obj)  # msg is a bytes object

Retrieval of the object is as follows:

import umsgpack
# Retrieve the message msg
obj = umsgpack.dumps(msg)

An ingenious feature of the standard is its extensibility. This can be used to add support for additional Python types or user defined classes. This example shows complex data being supported as if it were a native type:

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

Reading back:

import umsgpack
import umsgpack_ext  # Decoder only needs access to this module
with open('data', 'rb') as f:
    z = umsgpack.load(f)
print(z)  # z is complex

Please see this repo. The docs include references to the standard and to other implementations. The repo includes an asynchronous receiver which enables incoming messages to be decoded as they arrive while allowing other tasks to run concurrently.

5. Protocol Buffers

This is a Google standard described in this Wikipedia article. The aim is to provide a language independent, efficient, binary data interface. Records are variable length, and strings and integers of arbitrary size may be accommodated. The implementation compatible with MicroPython is a "micro" implementation: .proto files are not supported. However the data format aims to be a subset of the Google standard and claims compatibility with other platforms and languages.

The principal benefit to developers using only CPython/MicroPython is its efficient support for fields whose length varies at runtime. To my knowledge it is the sole solution for encoding such data in a compact binary format.

The following notes should be read in conjunction with the official docs. The notes aim to reduce the learning curve which I found a little challenging.

In normal use the object transmitted by minipb will be a dict with entries having various predefined data types. Entries may be objects of variable length including strings, lists and other dict instances. The structure of the dict is defined using a schema. Sender and receiver share the schema with each script using it to instantiate the Wire class. The Wire instance is then repeatedly invoked to encode or decode the data.

The schema is a tuple defining the structure of the data dict. Each element declares a key and its data type in an inner tuple. Elements of this inner tuple are strings, with element 0 defining the field's key. Subsequent elements define the field's data type; in most cases the data type is defined by a single string.

5.1 Installation

The library comprises a single file minipb.py. It has some dependencies, the logging module logging.py and bisect module bisect.py which may be found in micropython-lib. On RAM constrained platforms minipb.py may be cross-compiled or frozen as bytecode for even lower RAM consumption.

5.2 Data types

These are listed in the docs. Many of these are intended to maximise compatibility with the native data types of other languages. Where data will only be accessed by CPython or MicroPython, a subset may be used which maps onto Python data types:

1. 'U' A UTF8 encoded string.
2. 'a' A bytes object.
3. 'b' A bool.
4. 'f' A float A 32-bit float: the usual MicroPython default.
5. 'z' An int: a signed arbitrary length integer. Efficiently encoded with an ingenious algorithm.
6. 'd' A double precision 64-bit float. The default on Pyboard D SF6. Also on other platforms with special firmware builds.
7. 'x' An empty field.

5.2.1 Required and Optional fields

If a field is prefixed with * it is a required field, otherwise it is optional. The field must still exist in the data: the only difference is that a required field cannot be set to None. Optional fields can be useful, notably for boolean types which can then represent three states.

5.3 Application design

The following is a minimal example which can be pasted at the REPL:

import minipb

schema = (('value', 'z'),)  # Dict will hold a single integer
w = minipb.Wire(schema)

data = {'value': 0}
data['value'] = 150
tx = w.encode(data)
rx = w.decode(tx)  # received data
print(rx)

This example glosses over the fact that in a real application the data will change and the length of the transmitted string tx will vary. The receiving process needs to know the length of each string. Note that a consequence of the binary format is that delimiters cannot be used. The length of each record must be established and made available to the receiver. In the case where data is being saved to a binary file, the file will need an index. Where data is to be transmitted over and interface each string should be prepended with a fixed length "size" field. The following example illustrates this.

5.4 Transmitter/Receiver example

These examples can't be cut and pasted at the REPL as they assume send(n) and receive(n) functions which access the interface.

Sender example:

import minipb
schema = (('value', 'z'),
          ('float', 'f'),
          ('signed', 'z'),)
w = minipb.Wire(schema)
# Create a dict to hold the data
data = {'value': 0,
        'float': 0.0,
        'signed' : 0,}
while True:
    # Update values then encode and transmit them, e.g.
    # data['signed'] = get_signed_value()
    tx = w.encode(data)
    # Data lengths may change on each iteration
    # here we encode the length in a single byte
    dlen = len(tx).to_bytes(1, 'little')
    send(dlen)
    send(tx)

Receiver example:

import minipb
# schema must match transmitter. Typically both would import this.
schema = (('value', 'z'),
          ('float', 'f'),
          ('signed', 'z'),)

w = minipb.Wire(schema)
while True:
    dlen = receive(1)  # Data length stored in 1 byte
    data = receive(dlen)  # Retrieve actual data
    rx = w.decode(data)
    # Do something with the received dict

5.5 Repeating fields

This feature enables variable length lists to be encoded. List elements must all be of the same (declared) data type. In this example the value and txt fields are variable length lists denoted by the '+' prefix. The value field holds a list of int values and txt holds strings:

import minipb
schema = (('value', '+z'),
          ('float', 'f'),
          ('txt', '+U'),
          )
w = minipb.Wire(schema)

data = {'value': [150, 123, 456],
        'float': 1.23,
        'txt' : ['abc', 'def', 'ghi'],
        }
tx = w.encode(data)
rx = w.decode(tx)
print(rx)
data['txt'][1] = 'the quick brown fox'  # Strings have variable length
data['txt'].append('the end')  # List has variable length
data['value'].append(999)  # Variable length
tx = w.encode(data)
rx = w.decode(tx)
print(rx)

5.5.1 Packed repeating fields

This feature reduces some space overhead of encoded message caused by repeatedly emitting field headers as seen in regular repeated fields.

>>> import minipb
>>> normal=minipb.Wire('+z')
>>> len(normal.encode(range(10000)))
31744
>>> packed=minipb.Wire('#z')
>>> len(packed.encode(range(10000)))
21748
>>>

The author of minipb does not recommend their use for strings, bytes and nested messages due to compatibility concerns with the official Google Protobuf standards, which disallows such use.

5.6 Message fields (nested dicts)

The concept of message fields is a Protocol Buffer notion. In MicroPython terminology a message field contains a dict whose contents are defined by another schema. This enables nested dictionaries whose entries may be any valid protobuf data type.

This is illustrated below. The example extends this by making the field a variable length list of dict objects (with the '+[' specifier):

import minipb
# Schema for the nested dictionary instances
nested_schema = (('str2', 'U'),
                 ('num2', 'z'),)
# Outer schema
schema = (('number', 'z'),
          ('string', 'U'),
          ('nested', '+[', nested_schema),
          ('num', 'z'),)
w = minipb.Wire(schema)

data = {
    'number': 123,
    'string': 'test',
    'nested': [{'str2': 'string','num2': 888,},
               {'str2': 'another_string', 'num2': 12345,}, ],
    'num' : 42
}
tx = w.encode(data)
rx = w.decode(tx)
print(rx)
print(rx['nested'][0]['str2'])  # Access inner dict instances
print(rx['nested'][1]['str2'])
# Appending to the nested list of dicts
data['nested'].append({'str2': 'rats', 'num2':999})
tx = w.encode(data)
rx = w.decode(tx)
print(rx)
print(rx['nested'][2]['str2'])  # Access inner dict instances

5.6.1 Recursion

This is surely overkill in most MicroPython applications, but for the sake of completeness message fields can be recursive:

import minipb
inner_schema = (('str2', 'U'),
                ('num2', 'z'),)

nested_schema = (('inner', '+[', inner_schema),)

schema = (('number', 'z'),
          ('string', 'U'),
          ('nested', '[', nested_schema),
          ('num', 'z'),)

w = minipb.Wire(schema)

data = {
       'number': 123,
       'string': 'test',
       'nested': {'inner':({'str2': 'string', 'num2': 888,},
                           {'str2': 'another_string','num2': 12345,}, ),},
        'num' : 42
        }
tx = w.encode(data)
rx = w.decode(tx)
print(rx)
print(rx['nested']['inner'][0]['str2'])