instruct 0.7.5.post2

Compact/flexible/fast/strictly-typed/(add other ridiculous demands) object system - similar to Pydantic but focused on `__slot__`ed objects

Issues	https://github.com/autumnjolitz/instruct/issues
Source	https://github.com/autumnjolitz/instruct

A compact, fast object system that can serve as the basis for a DAO model.

To that end, instruct uses __slots__ to prevent new attribute addition, properties to control types, event listeners and historical changes, and a Jinja2-driven codegen to keep a pure-Python implementation as fast and as light as possible.

I want to basically have a form of strictly typed objects that behave like C structs but can handle automatically coercing incoming values correctly, have primitive events and have fast __iter__, __eq__ while also allowing for one to override it in the final class (and even call super!)

This girl asks for a lot but I like taking metaclassing as far as it can go without diving into using macropy. 😉

Current Capabilities:

• Support multiple inheritance, chained fields and __slots__ [Done]

• Support type coercions (via _coerce__) [Done]

• Strictly-typed ability to define fixed data objects [Done]

• Ability to drop all of the above type checks [Done]

• Track changes made to the object as well as reset [Done]

• Fast __iter__ [Done]

• Native support of pickle [Done]/json [Done]

• Support List[type] declarations and initializations [Done]

• optionally data class annotation-like behavior [Done]

• _asdict, _astuple, _aslist functions like in a NamedTuple [Done]

• get, keys, values, item functions available in the module and in a mixin named mapping=True

  • This effectively allows access like other packages e.g. attrs.keys(item_instance)

• bytes/bytearray are urlsafe base64 encoded by default, can override per field via a class level BINARY_JSON_ENCODERS = {key: encoding_function} [Done]

• Allow __coerce__ to have a tuple of field names to avoid repetition on __coerce__ definitions [Done]

• Allow use of Literal in the type (exact match of a value to a vector of values) [Done]

• Allow subtraction of properties like (F - {"a", "b"}).keys() == F_without_a_b.keys() [Done] + This will allow one to slim down a class to a restricted subtype, like for use in a DAO system to load/hold less data.

• Allow subtraction of properties like (F - {"a": {"b"}).keys() == F_a_without_b.keys() [Done] + This allows for one to remove fields that are unused prior to class initialization.

• Allow subtraction of properties via an inclusive list like (F & {"a", "b"}).keys() == F_with_only_a_and_b.keys() [Done]

• Allow subtraction to propagate to embedded Instruct classes like (F - {"a.b", "a.c"}).a.keys() == (F_a.keys() - {"b", "c")) [Done] + This would really allow for complex trees of properties to be rendered down to thin SQL column selects, thus reducing data load.

• Replace references to an embedded class in a __coerce__ function with the subtracted form in case of embedded property subtractions [Done]

• Allow use of Annotated i.e. field: Annotated[int, NoJSON, NoPickle] and have to_json and pickle.dumps(...) skip “field” [Done] + Would grant a more powerful interface to controlling code-gen’ed areas via cls._annotated_metadata (maps field -> what’s inside the Annotation) [Done]

Next Goals:

• Allow Generics i.e. class F(instruct.Base, T): ... -> F[str](...) + Would be able to allow specialized subtypes

• CStruct-Base class that operates on an _cvalue cffi struct.

• Cython compatibility

Design Goal

This comes out of my experience of doing multiple object systems mean to represent database relations and business rules. One thing that has proven an issue is the requirements for using as little memory as possible, as little CPU as possible yet prevent the developer from trying to stick a string where a integer belongs.

Further complicating this model is that desire to “correct” data as it comes in. Done correctly, it is possible to feed an instruct.Base-derived class fields that are not of the correct data type but are eligible for being coerced (converted) into the right type with a function. With some work, it’ll be possible to inline a lambda val: ... expression directly into the setter function code.

Finally, multiple inheritance is a must. Sooner or later, you end up making a single source implementation for a common behavior shared between objects. Being able to share business logic between related implementations is a wonderful thing.

Wouldn’t it be nice to define a heirachy like this:

import pickle
import datetime
from typing import List

try:
    from typing import Annotated
except ImportError:
    from typing_extensions import Annotated
from instruct import Base, NoJSON, NoIterable, NoPickle, NoHistory


class Member(Base):
    __slots__ = {"first_name": str, "last_name": str, "id": int}

    def _set_defaults(self):
        self.first_name = self.last_name = ""
        self.id = -1
        super()._set_defaults()


class Organization(Base, history=True):
    # ARJ: Note how we can also use the dataclass/typing.NamedTuple
    # definition format and it behaves just like the ``__slots__`` example
    # above!
    name: str
    id: int
    members: List[Member]
    created_date: datetime.datetime
    secret: Annotated[str, NoJSON, NoPickle, NoIterable, NoHistory]

    __coerce__ = {
        "created_date": (str, lambda obj: datetime.datetime.strptime("%Y-%m-%d", obj)),
        "members": (List[dict], lambda values: [Member(**value) for value in values]),
    }

    def _set_defaults(self):
        self.name = ""
        self.id = -1
        self.members = []
        self.created_date = datetime.datetime.utcnow()
        super()._set_defaults()

And have it work like this?

data = {
    "name": "An Org",
    "id": 123,
    "members": [{"id": 551, "first_name": "Jinja", "last_name": "Ninja"}],
}
org = Organization(secret="my secret", **data)
assert org.members[0].first_name == "Jinja"
assert org.secret == "my secret"
org.name = "New Name"
org.created_date = datetime.datetime(2018, 10, 23)
print(tuple(org.list_changes()))
# Returns
# (
#     LoggedDelta(timestamp=1652412832.7408261, key='name', delta=Delta(state='default', old=Undefined, new='', index=0)),
#     LoggedDelta(timestamp=1652412832.7408261, key='id', delta=Delta(state='default', old=Undefined, new=-1, index=0)),
#     LoggedDelta(timestamp=1652412832.7408261, key='members', delta=Delta(state='default', old=Undefined, new=[], index=0)),
#     LoggedDelta(timestamp=1652412832.7408261, key='created_date', delta=Delta(state='default', old=Undefined, new=datetime.datetime(2022, 5, 13, 3, 33, 52, 740650), index=0)),
#     LoggedDelta(timestamp=1652412832.740923, key='id', delta=Delta(state='initialized', old=-1, new=123, index=4)),
#     LoggedDelta(timestamp=1652412832.741002, key='members', delta=Delta(state='initialized', old=[], new=[<__main__.Member._Member object at 0x104364640>], index=5)),
#     LoggedDelta(timestamp=1652412832.741009, key='name', delta=Delta(state='initialized', old='', new='An Org', index=6)),
#     LoggedDelta(timestamp=1652412832.741021, key='name', delta=Delta(state='update', old='An Org', new='New Name', index=7)),
#     LoggedDelta(timestamp=1652412832.741031, key='created_date', delta=Delta(state='update', old=datetime.datetime(2022, 5, 13, 3, 33, 52, 740650), new=datetime.datetime(2018, 10, 23, 0, 0), index=8))
# )

assert not any(y == "my secret" for y in tuple(org))
assert Organization.to_json(org) == {
    "created_date": "2018-10-23T00:00:00",
    "id": 123,
    "members": [{"first_name": "Jinja", "id": 551, "last_name": "Ninja"}],
    "name": "New Name",
}
org2 = pickle.loads(pickle.dumps(org))
assert org2.secret is None
assert org2.to_json() == {
    "created_date": "2018-10-23T00:00:00",
    "id": 123,
    "members": [{"first_name": "Jinja", "id": 551, "last_name": "Ninja"}],
    "name": "New Name",
}

Example Usage

>>> from instruct import Base
>>>
>>> class MyClass(Base):
...     foo: int
...     bar: Optional[str]
...     baz: Union[Dict[str, str], int]
...     def __eq__(self, other):
...         if isinstance(other, tuple) and len(other) == 3:
...            # Cast the tuple to this type!
...            other = MyClass(*other)
...         return super().__eq__(other)
...
>>> instance = MyClass(1, None, baz={"a": "a"})
>>> assert instance.foo == 1
>>> assert instance.bar is None
>>> instance.bar = "A String!"
>>>
>>> assert instance == (1, "A String!", {"a": "a"})
>>>
>>> instance.foo = 'I should not be allowed'
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "<getter-setter>", line 36, in _set_foo
TypeError: Unable to set foo to 'I should not be allowed' (str). foo expects a int
>>>

Instruct adds a Range type for use in Annotated[...] type definitions.

Range

class Range(lower, upper, flags: RangeFlags = <RangeFlags.CLOSED_OPEN: 4>, *, type_restrictions: Tuple[Type, ...]=())
    ...

lower and upper can be anything that supports __lt__, __gt__, __eq__.

type_restrictions can be used to apply a Range constraint to some value types.

flags can be used to set the interval type. Default is closed-open [).

>>> from typing import Tuple, type
>>> from instruct import Range, RangeFlags, RangeError
>>> lower, upper = 0, 255
>>> r = Range(lower, upper, flags: RangeFlags = RangeFlags.CLOSED_OPEN)
>>> 10 in r
True
>>> 0 in r
True
>>> 256 in r
False

When used inside an instruct-derived class, an attempt to assign a value that doesn’t satisfy a tuple of ranges will throw a RangeError (inherits from ValueError and TypeError).

Inside is the value (what was rejected) and a copy of the ranges at ranges that were tried (and failed). If the type_restrictions are specified in a range, it will not be tried if the value type isn’t applicable.

class RangeError(value: Any, ranges: Tuple[Range, ...], message: str="")
    ...

Example:

>>> from instruct import SimpleBase, Range
>>> from typing_extensions import Annotated
>>> from typing import Union
>>> class Planet(SimpleBase):
...     mass_kg: Annotated[Union[float, int], Range(600 * (10**18), 1.899e27)]
...     radius_km: Annotated[Union[float, int], Range(2439.766, 142_800)]
...
>>>
>>> mercury = Planet(3.285 * (10**23), 2439.766)
>>> mars = Planet(0.64169 * (10**24), 3376.2)
>>>
>>> pluto = Planet(1.30900 * (10**22), 1188.30742)
Traceback (most recent call last):
  File "/Users/autumn/software/instruct/instruct/__init__.py", line 2113, in __init__
    setattr(self, key, value)
  File "<getter-setter>", line 30, in _set_radius_km
  File "/Users/autumn/software/instruct/instruct/typedef.py", line 40, in __instancecheck__
    return func(instance)
  File "/Users/autumn/software/instruct/instruct/typedef.py", line 227, in test_func
    raise RangeError(value, failed_ranges)
instruct.exceptions.RangeError: ('Unable to fit 1188.30742 into [2439.766, 142800)', 1188.30742, (Range(2439.766, 142800, flags=CLOSED_OPEN, type_restrictions=()),))

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/autumn/software/instruct/instruct/__init__.py", line 2128, in __init__
    self._handle_init_errors(errors, errored_keys, unrecognized_keys)
  File "/Users/autumn/software/instruct/instruct/__init__.py", line 2094, in _handle_init_errors
    ) from errors[0]
instruct.exceptions.ClassCreationFailed: ('Unable to construct Planet, encountered 1 error', RangeError('Unable to fit 1188.30742 into [2439.766, 142800)', 1188.30742, (Range(2439.766, 142800, flags=CLOSED_OPEN, type_restrictions=()),)))
>>>

Comparison to Pydantic

Pydantic is a much larger project with many more eyes. Instruct was designed from the beginning to support multiple-inheritance and __slot__ specialization. Pydantic does much the same as Instruct. Pydantic is much more feature-filled and infinitely more popular. Instruct is a one-woman crew.

Instruct was a reflexive response to years of dealing with needing to handle Object-Relational impedance mismatch in MySQL/Postgres. It was meant as a building block for enabling templated SQL writing in a controlled manner without resorting to ORMs (more akin to DAO approach). As such, its design and evolution reflects that.

Instruct is not better. Nor is it worse. Instruct simply does what it’s designed to do and no more.

I suggest you use Pydantic if you’re interested in a far bigger, far more lively, far better supported library. Instruct has different ambitions and does not intend to replace or compete with Pydantic.

Instruct was designed in October 7, 2017 but was released in Dec 9, 2018.

Pydantic’s earliest release (0.1.0) is in 2017-06-03.

Design differences between the two:

• Instruct attempts to NOT provide functions/attributes that may be clobbered via SimpleBase and remapping the public variables to _{{varname}}_

  • Pydantic allows one to override the remapping, but does occupy names like dict, json, etc,.

• Pydantic provides Model properties like dict(), json(), copy(), etc

  • Instruct Base (via JSONSerializable) provides to_json, __json__, from_json, from_many_json

  • If you use SimpleBase, you can access similar properties ONLY on the class itself (we do not attach it to the class instance to avoid clobbering)

• Instruct is shifting to a paradigm of using free-functions like asdict, astuple, keys, items, values, etc instead of clobbering fields on an object

  • we want to allow as many user-specified names as possible

• Instruct wants to remain small

• Instruct wants to support CStruct``s and possible basis for using a ``bytearray as the underlying memory for enabling rich types while allowing a near memcpy.

Things Instruct can do that Pydantic doesn’t:

• Class subtraction and masking

  • You can subtract out a field by a string represetation, multiple by subtracting out an Iterable[str], or even apply such via a nested dict (where the values are None or another mapping to apply to a sub-object)

  • You can cls & {"field"} or cls & {"field": {"keep_this"}} and get a class with only field and field.keep_this

• Allows unsupported types by fields to call functions to parse/coerce it into a valid value (__coerce__)

  • Pydantic suggests you use Data bind to handle weirdies

  • Pydantic does a lot of conversions for you automatically

  • Instruct demands you make them explicit in your handling functions.

• Instruct creates custom types representing complex, nested data structures such it does an effect isinstance(value, ComplexType) to verify if a complex, nested tree of objects does match.

  • The types are meant only for an isinstance check.

Things Pydantic does that Instruct doesn’t:

• Discriminated Unions (Current approach in Instruct is to add the common class into the Union and specialize after __init__ or do it in the __coerce__ phase)

• Type/Callable/Generator attribute assignment

• Generics (on my todo)

• validation (instruct is used to provide the building blocks for validation, not doing it by itself. That might change.)

• actual mypy, vscode, pycharm, etc integration

• schema export

• aliases (Instruct expects you to just add a @property that gets/sets the true field)

• lots more little goodies

Design

Solving the multiple-inheritance and __slots__ problem

Consider the following graph:

Base1    Base2
     \  /
   Class A

If both defined __slots__ = (), Class A would be able to declare __slots__ to hold variables. For now on, we shall consider both Base’s to have __slots__ = () for simplicity.

However, consider this case:

Base1    Base2
     \  /
   Class A     Class B
          \    /
          Class C

Now this isn’t possible if Class A has non-empty __slots__.

But what if we could change the rules. What if, somehow, when you __new__ ed a class, it really gave you a specialized form of the class with non-empty __slots__?

Such a graph may look like this:

Base1    Base2
     \  /
   Class A     Class B
      |  \    /     |
Class _A  Class C  Class _B
            |
          Class _C

Now it is possible for any valid multiple-inheritance chain to proceed, provided it respects the above constraints - there are either support classes or data classes (denoted with an underscore in front of their class name). Support classes may be inherited from, data classes cannot.

Solving the Slowness issue

I’ve noticed that there are constant patterns of writing setters/getters and other related functions. Using Jinja2, we can rely on unhygenic macros while preserving some semblance of approachability. It’s more likely a less experienced developer could handle blocks of Jinja-fied Python than AST synthesis/traversal.

Callgraph Performance

Release Process

$ rm -rf dist/* && python -m pytest tests/ && python setup.py sdist bdist_wheel && twine upload dist/*

Benchmark

Latest benchmark run::

(python) Fateweaver:~/software/instruct [master]$ python --version
Python 3.7.7
(python) Fateweaver:~/software/instruct [master]$ python -m instruct benchmark
Overhead of allocation, one field, safeties on: 19.53us
Overhead of allocation, one field, safeties off: 19.50us
Overhead of setting a field:
Test with safeties: 0.27 us
Test without safeties: 0.17 us
Overhead of clearing/setting
Test with safeties: 0.75 us
Test without safeties: 0.65 us
(python) Fateweaver:~/software/instruct [master]$

Before additions of coercion, event-listeners, multiple-inheritance

$ python -m instruct benchmark
Overhead of allocation, one field, safeties on: 6.52us
Overhead of allocation, one field, safeties off: 6.13us
Overhead of setting a field:
Test with safeties: 0.40 us
Test without safeties: 0.22 us
Overhead of clearing/setting
Test with safeties: 1.34 us
Test without safeties: 1.25 us

After additions of those. Safety is expensive.

$ python -m instruct benchmark
Overhead of allocation, one field, safeties on: 19.25us
Overhead of allocation, one field, safeties off: 18.98us
Overhead of setting a field:
Test with safeties: 0.36 us
Test without safeties: 0.22 us
Overhead of clearing/setting
Test with safeties: 1.29 us
Test without safeties: 1.14 us