MSB Architecture

Mega-Super-Base (MSB) is an architecture for Python applications built around a single entry point. You describe your data as typed entities, you describe what may be done to them as operations, and everything reaches both through one orchestrator — a user, a GUI, another API, whatever drives the application.

A request is data, not a call:

{"operation": "configure", "obj": telescope, "attributes": {"set_diameter": 64.0}}

which is what lets the same code serve a dialog box, a script and a remote caller, and what lets a session be logged and replayed.

Features

• Typed entities: attributes validated against their annotations, nested to any depth, including List, Dict, Tuple, Set, Union, Literal, Callable and Type[X].
• Constraints on values, not just types: price: Annotated[float, Positive()] is enforced on construction, on assignment and on restore, with no __init__ of your own.
• Containers for collections: named, queryable, serializable, with bulk operations.
• One entry point: a Manipulator registers operations and processes requests; the per-operation facades are sugar so you rarely write a request dictionary by hand.
• Reading and writing come free: inspect and configure are supplied, so an application that only reads and writes its model needs no operation layer at all.
• Operations that write themselves: a handler is usually one call to _apply_methods, which applies everything a request names and reports each outcome.
• Serialization that round-trips: json.loads(json.dumps(obj.to_dict())) restores an equal object, through lists, dicts, sets and tuples, for entities nested to any depth. Cycles are detected rather than followed, and serialized data carries the version of the class that wrote it, so a model can change shape and still read its old files.
• Logging that behaves: a dedicated msb_arch logger that stays silent until the application configures it.
• Exceptions you can catch precisely: everything derives from MSBError, and also from the built-in it replaces, so except TypeError keeps working while except DuplicateNameError becomes possible.
• One place to hang metrics, auditing, rate limiting and authorisation: an interceptor sees a request before it runs and its response after, and may refuse or rewrite it. Request metrics and a replayable request journal ship using nothing more than that hook.
• Asynchronous when you need it: await manipulator.ainspect(...) keeps an event loop responsive by moving the work off it, and every synchronous signature is untouched.
• No external dependencies: Python >= 3.12 and nothing else.

Installation

pip install msb_arch

Quick Start

Describe the data, describe the operations, drive both through the orchestrator.

from msb_arch import BaseContainer, BaseEntity, Manipulator

# 1. the data
class Telescope(BaseEntity):
    diameter: float

    def get_diameter(self) -> float:
        return self.diameter

    def set_diameter(self, value: float) -> bool:
        self.diameter = value
        return True

class Telescopes(BaseContainer[Telescope]):
    pass

# 2. the entry point
class Observatory(Manipulator):
    pass

manipulator = Observatory(base_classes=[Telescope, Telescopes])

dishes = Telescopes(name="array")
dishes.add(Telescope(name="DSS14", diameter=70.0))
dish = dishes.get("DSS14")

manipulator.inspect(dish, get_diameter=None)      # 70.0
manipulator.configure(dish, set_diameter=64.0)
manipulator.inspect(dish, get_diameter=None)      # 64.0

dishes.to_dict()["items"]["DSS14"]["diameter"]    # 64.0

There is no operation layer to write: inspect and configure follow from the request model itself, so the framework supplies them, and they serve every type. You write a Super when an operation carries domain logic — calculate, visualize — and register it the same way.

from msb_arch import Super

class Calculator(Super):
    OPERATION = "calculate"

    def _calculate_telescope(self, obj, attributes):
        return self._apply_methods(obj, attributes)

manipulator.register_operation(Calculator(manipulator))

A handler is one line because _apply_methods owns the loop, and the orchestrator dispatches by operation and by the type of the object, so adding an entity adds no code at all.

Ask for several things at once and every outcome comes back, whatever the order:

manipulator.inspect(dish, get_diameter=None, get=["name", "isactive"])
# {'get_diameter': {'status': True, 'result': 64.0},
#  'get':          {'status': True, 'result': {'name': 'DSS14', 'isactive': True}}}

Run several requests as one batch:

manipulator.batch([
    {"operation": "configure", "obj": dish, "attributes": {"set_diameter": 70.0}},
    {"operation": "inspect", "obj": dish, "attributes": {"get_diameter": None}},
])

Architecture

Four modules, three layers.

Main classes:

• Serializable — what an entity and a container have in common: annotated fields and their validation, name and isactive, to_dict, the cache. Use it in isinstance checks that should accept either.
• BaseEntity — an object addressed by its attributes.
• BaseContainer[T] — a named collection addressed by its items. A sibling of BaseEntity, not a subclass: an entity and a container mean different things by get, set and clear.
• Super — an operation. Subclass it, name the operation, and write handlers as _<operation>_<type> or _<operation> for the fallback.
• Project — a named collection of entities with a factory for creating them.
• Manipulator — the entry point. Registers operations, processes requests, generates a facade per operation.
• MethodResults — what an operation reports: every method it ran, mapped to its outcome.

Documentation

• Guide — start here: a working application, built from nothing
• API reference — every public class and method
• Compatibility — what will not break, and how anything changes
• Architecture and diagrams
• Base module — the data model, type hints, serialization, caching
• Super module — writing your own operation
• Mega module — interceptors, built-ins, the asynchronous surface
• Examples
• Roadmap — what comes after 1.0
• Changelog — release history and upgrade notes

Testing

Unit, integration, performance and concurrency suites, run with pytest.

The tests import msb_arch rather than the source tree, so they exercise whatever is installed. Install the package first:

pip install -e .

Then run them:

pytest tests/

CI builds the wheel, installs it, checks that msb_arch resolves inside site-packages, and runs the same suites against it — so the distribution that ships is the one that was tested.

License

MSB is licensed under the MSB Software License for non-commercial and research use, allowing free use, modification, and distribution for non-commercial purposes with attribution.

For commercial use, a separate royalty-bearing license is required. Please contact almax1024@gmail.com for details.

Contacts

• Author: Alexey Rudnitskiy
• Email: almax1024@gmail.com
• Repository: https://github.com/Torward1024/MSB
• Version: 1.1.1