United Income at Capital One created this project with the intention of it helping us with JSON parsing. This project has not gained wide adoption. As such, we have stopped providing updates to this project and archived it as of June 28th, 2021.

json-syntax

A Python library to translate between JSON compatible structures and native Python classes using customizable rules.

Use case

If you're like the authors, you tried writing an encoding function that attempted to encode and decode by interrogating the types at runtime, maybe calling some method like asdict. This works fine for generating JSON, but it gets sketchy¹ when trying to decode the same JSON.

Further, we have annotations in Python 3! Even if you're not using a type checker, just labeling the types of fields makes complex data structures far more comprehensible.

This library is aimed at projects that have a complex JSON schema that they're trying to structure using libraries like attrs.

It exploits gradual typing via annotations, typing and dataclasses
It expects classes to be statically described using types
- But a fallback can be provided to handle data described at runtime
- It provides hooks to normalize legacy inputs
It makes it trivial to extend the library with your own rules
- Actions and Rules are simply functions
- Encoders and decoders can be pickled
The library has no dependencies of its own on python 3.7+
- It does not read or write JSON

Supported types

Atoms including None, bool, int, float, str.
- Floats may optionally be represented as strings.
The decimal.Decimal class, represented as itself or in string form.
The datetime.date and datetime.datetime classes, represented in ISO8601 form.
Preliminary support for datetime.timedelta as ISO8601 time durations.
Subclasses of enum.Enum, represented by the string names.
- Also, a faux_enums rule will accept an Enum type if you just use strings in your code.
The typing.Optional[E] type allows a JSON null to be substituted for a value.
Collections including typing.List[E], typing.Tuple[E, ...], typing.Set[E] and typing.FrozenSet[E].
- The ... is literal and indicates a homogenous tuple, essentially a frozen list.
The typing.Dict[K, V] type allows a JSON object to represent a homogenous dict.
- Restriction: the keys must be strings, ints, enums or dates.
New: The typing.TypedDict type allows a JSON object to represent a dict with specific keys.
Python classes implemented using attrs.attrs, dataclasses.dataclass are represented as JSON dicts and
Named tuples via typing.NamedTuple and heterogenous tuples via typing.Tuple.
- Though, you should consider converting these to dataclass.
The typing.Union[A, B, C] rule will recognize alternate types by inspection.

In addition, dataclass and attrs classes support hooks to let you completely customize their JSON representation.

Extras

These were originally intended as examples for how to use the package, but they're potentially useful in their own right.

A ruleset for use with AWS DynamoDB is included with basic facilities.
- Restriction: No general support for typing.Union, only Optional.
- Restriction: No general support for Set, only the special cases that are native to DynamoDB.
A Flag pseudo-type allows you to use regular strings directly as flags.
A rule that will accept a complete datetime and return a date by truncating the timestamp.

Usage

This example is also implemented in unit tests. First, let's declare some classes.

import json_syntax as syn
from dataclasses import dataclass  # attrs works too
from decimal import Decimal
from datetime import date
from enum import Enum

@dataclass
class Account:
    user: str
    transactions: List['Trans']  # Forward references work!
    balance: Decimal = Decimal()

class TransType(Enum):
    withdraw = 0
    deposit = 1

@dataclass
class Trans:
    type: TransType
    amount: Decimal
    stamp: date

We'll next set up a RuleSet and use it to construct an encoder. The std_ruleset function is a one-liner with some reasonable overrides. Here, we've decided that because some intermediate services don't reliably retain decimal values, we're going to represent them in JSON as strings.

>>> rules = syn.std_ruleset(decimals=syn.decimals_as_str)
>>> encode_account = rules.python_to_json(typ=Account)
>>> decode_account = rules.json_to_python(typ=Account)

The RuleSet examines the type and verb, searches its list of Rules, and then uses the first one that handles that type and verb to produce an Action.

For example, attrs_classes is a Rule that recognizes the verbs python_to_json and json_to_python and will accept any class decorated with @attr.s or @dataclass.

It will scan the fields and ask the RuleSet how to encode them. So when it sees Account.user, the atoms rule will match and report that converting a str to JSON can be accomplished by simply calling str on it. The action it returns will literally be the str builtin.

Thus attrs_classes will build a list of attributes on Account and actions to convert them, and constructs an action to represent them.

>>> sample_value = Account(
...     'bob', [
...         Trans(TransType.withdraw, Decimal('523.33'), date(2019, 4, 4))
...     ], Decimal('77.00')
... )

>>> encode_account(sample_value)
{
  'user': 'bob',
  'transactions': [
    {
      'type': 'withdraw',
      'amount': '523.33',
      'stamp': '2019-04-04'
    }
  ], 'balance': '77.00'
}

Encoding and decoding

The aim of all this is to enable reliable usage with your preferred JSON library:

with open('myfile.json', 'r') as fh:
    my_account = decode_account(json.load(fh))

with open('myfile.json', 'w') as fh:
    json.dump(encode_account(my_account))

Using generic types

Generally, the typing module simply provides capital letter type names that explicitly correspond to the internal types. See TYPES for a more thorough introduction.

And you specify the type of the contents as a parameter in square brackets.

Thus we have:

list and List[E]
set and Set[E]
tuple and Tuple[E, ...] is a special case!
frozenset and FrozenSet[E]
dict and Dict[K, V]

Tuple is a special case. In Python, they're often used to mean "frozenlist", so Tuple[E, ...] (the ... is the Ellipsis object) indicates all elements have the type E.

They're also used to represent an unnamed record. In this case, you can use Tuple[A, B, C, D] or however many types. It's generally better to use a dataclass.

The standard rules don't support:

Using abstract types like Iterable or Mapping.
Using type variables.
Any kind of callable, coroutine, file handle, etc.

Support for deriving from Generic

There is experimental support for deriving from typing.Generic. An attrs or dataclass may declare itself a generic class. If another class invokes it as YourGeneric[Param, Param], those Param types will be substituted into the fields during encoding. This is useful to construct parameterized container types. Example:

@attr.s(auto_attribs=True)
class Wrapper(Generic[T, M]):
    body: T
    count: int
    messages: List[M]

@attr.s(auto_attribs=True)
class Message:
    first: Wrapper[str, str]
    second: Wrapper[Dict[str, str], int]

Unions

A union type lets you present alternate types that the converters will attempt in sequence, e.g. typing.Union[MyType, int, MyEnum].

This is implemented in the unions rule as a so-called² undiscriminated union. It means the module won't add any additional information to the value such as some kind of explicit tag.

When converting from Python to JSON, the checks are generally just using isinstance, but when converting from JSON to Python, the check may be examining strings and dict fields.

Thus, ambiguous values, especially JSON representations, may confuse the decoder. See the section on sharp edges for more details.

Hooks

We'll first examine decode and encode hooks. These let us entirely rewrite the JSON representation before the normal logic is applied.

Let's suppose our Account class used to name the balance field bal and we need to support legacy users.

@dataclass
class Account:
    @classmethod
    def __json_pre_decode__(cls, value):
        if 'bal' in value:
            value = dict(value)
            value['balance'] = value.pop('bal')
        return value

    def __json_post_encode__(self, value):
        return dict(value, bal=value['balance'])

    ...

When we decode the value, the following sequence of steps takes place:

__json_pre_decode__ is called with {'user': 'bob', 'bal': '77.0', ...} and it returns {'user': 'bob', 'balance': '77.0', ...}
Decoders are called against user and balance and the other fields
The resulting dictionary is passed to Account(**result) to construct the instance.

During encoding, the reverse sequence takes place:

The instance's fields are read and passed to encoders.
The values are combined into a dict.
__json_post_encode__ is called with {'user': 'bob', 'balance': '77.0', ...} and can adjust the field name to bal.

JSON type check hook

Type checks are only used in json-syntax to support typing.Union; in a nutshell, the unions rule will inspect some JSON to see which variant is present.

If a type-check hook is not defined, __json_pre_decode__ will be called before the standard check is completed. (The standard check attempts to determine if required fields are present and have the correct type.)

If you have information that can determine the type faster, a check hook can help.

Going back to our Account example, suppose we decide to support multiple account types through a special class field. This is faster and more robust.

class AbstractAccount:
    @classmethod
    def __json_check__(cls, value):
        return isinstance(value, dict) and value.get('class') == cls.__name__

@dataclass
class AccountA(AbstractAccount):
    ...

encode_account = rules.lookup(typ=Union[AccountA, AccountB, AccountC],
                              verb='python_to_json')

Adding custom rules

See the extras for details on custom rules, but generally a rule is just a function. Say, for instance, your type has class methods that encode and decode, this would be sufficient for many cases:

def my_rule(verb, typ, ctx):
    if issubclass(typ, MyType):
        if verb == 'json_to_python':
            return typ.decoder
        elif verb == 'python_to_json':
            return typ.encoder

If your rule needs an encoder or decoder for a standard type, it can call ctx.lookup(verb=verb, typ=subtype). The helper functions defined in json_syntax.action_v1 are intended to stay the same so that custom rules can reuse them.

Debugging ambiguous structures

(May need more docs and some test cases.)

As json-syntax tries to directly translate your Python types to JSON, it is possible to write ambiguous structures. To avoid this, there is a handy is_ambiguous method:

# This is true because both are represented as an array of numbers in JSON.
rules.is_ambiguous(typ=Union[List[int], Set[int]])

@dataclass
class Account:
    user: str
    address: str

# This is true because such a dictionary would always match the contents of the account.
rules.is_ambiguous(typ=Union[Dict[str, str], Account])

The aim of this is to let you put a check in your unit tests to make sure data can be reliably expressed given your particular case.

Internally, this is using the PATTERN verb to represent the JSON pattern, so this may be helpful in understanding how json-syntax is trying to represent your data:

print(rules.lookup(typ=MyAmbiguousClass, verb='show_pattern'))

Sharp edges

The RuleSet caches encoders. Construct a new ruleset if you want to change settings.

Encoders and decoders do very little checking. Especially, if you're translating Python to JSON, it's assumed that your Python classes are correct. The encoders and decoders may mask subtle issues as they are calling constructors like str and int for you. And, by design, if you're translating from JSON to Python, it's assumed you want to be tolerant of extra data.

Everything to do with typing. It's a bit magical and sort of wasn't designed for this. We have a guide to it to try and help.

Union types. You can use typing.Union to allow a member to be one of some number of alternates, but there are some caveats. You should use the .is_ambiguous() method of RuleSet to warn you of these.

Atom rules accept specific types. At present, the rules for atomic types (int, str, bool, date, float, Decimal) must be declared as exactly those types. With multiple inheritance, it's not clear which rule should apply

Checks are stricter than converters. For example, a check for int will check whether the value is an integer, whereas the converter simply calls int on it. Thus there are inputs for where MyType would pass but Union[MyType, Dummy] will fail. (Note that Optional is special-cased to look for None and doesn't have this problem.)

Numbers are hard. See the rules named floats, floats_nan_str, decimals, decimals_as_str for details on how to get numbers to transmit reliably. There is no rule for fractions or complex numbers as there's no canonical way to transmit them via JSON.

Maintenance

This package is maintained via the poetry tool. Some useful commands:

Setup: poetry install
Run tests: poetry run pytest tests/
Reformat: black json_syntax/ tests/
Generate setup.py: dephell deps convert -e setup
Generate requirements.txt: dephell deps convert -e req

Running tests via docker

The environments for 3.4 through 3.9 are in pyproject.toml, so just run:

dephell deps convert -e req  # Create requirements.txt
dephell docker run -e test34 pip install -r requirements.txt
dephell docker run -e test34 pytest tests/
dephell docker shell -e test34 pytest tests/
dephell docker destroy -e test34

Notes

1: Writing the encoder is deceptively easy because the instances in Python has complete information. The standard json module provides a hook to let you encode an object, and another hook to recognize dicts that have some special attribute. This can work quite well, but you'll have to encode all non-JSON types with dict-wrappers for the process to work in reverse. ↩

2: A discriminated union has a tag that identifies the variant, such as status codes that indicate success and a payload, or some error. Strictly, all unions must be discriminated in some way if different code paths are executed. In the unions rule, the discriminant is the class information in Python, and the structure of the JSON data. A less flattering description would be that this is a "poorly" discriminated union. ↩