MSB Architecture

Mega-Super-Base (MSB) is for the application that outgrew a script: one whose data has a shape worth validating, whose operations keep multiplying, and which now needs a window, a command line and a server over the same model without three copies of the logic.

You describe the data as typed entities, you describe what may be done to it as operations, and everything reaches both through one orchestrator. A request is data, not a call:

{"operation": "configure", "obj": part, "attributes": {"set": {"params": {"price": 4.5}}}}

Which is a small thing that buys several large ones:

• The same code serves a dialog, a script and a remote caller, because none of them is calling a method — they are all sending the same dictionary.
• A session can be recorded and replayed: the journal keeps each request as data, with the object's address in the model rather than a reference to it, so a recorded session runs again against a model built from scratch and lands on the objects it meant.
• Menus, dependency graphs, handler stubs and execution order are derived from the code rather than listed, so none of them can go stale.
• Metrics, auditing, authorisation and rate limiting hang on one hook that sees every request, and the operations know nothing about them.

Written for a working instrument — pAstroCORE, which plans radio astronomy observations — and everything here exists because that application needed it.

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; a facade per operation means you rarely write a request dictionary by hand.
• Reading and writing come free. inspect, configure, save, load and catalogue are registered for you, 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.
• Pipelines. Several requests that feed each other, given as data in one call. The order, what may run at once, and what to skip when a step fails all follow from the edges.
• Serialization that round-trips. json.loads(json.dumps(obj.to_dict())) restores an equal object, through lists, dicts, sets and tuples, nested to any depth. Cycles are detected rather than followed, and data carries the schema version of the class that wrote it.
• Derived answers instead of hand-written ones. What operations exist, what handlers they have, which handler needs which, which type holds which, and what a change reaches — all read back from the code, so a menu or a diagram cannot go stale.
• 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 journal ship using nothing more than that hook.
• A session that replays somewhere else. The journal records each request as plain data, with the object's address in the model rather than a reference to it, so a recorded session can be written to a file and run again against a model built from scratch -- and land on the objects it meant. address and locate are the same addressing on its own, for a request that has to cross a process or a wire.
• Asynchronous when you need it. await manipulator.ainspect(...) moves the work off the event loop, and every synchronous signature is untouched.
• 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.
• 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 Part(BaseEntity):
    price: float

    def get_price(self) -> float:
        return self.price

    def set_price(self, value: float) -> bool:
        self.price = value
        return True

class Parts(BaseContainer[Part]):
    pass

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

manipulator = Workshop(base_classes=[Part, Parts])

box = Parts(name="box")
box.add(Part(name="bolt", price=4.5))
bolt = box.get("bolt")

assert manipulator.inspect(bolt, get_price=None) == 4.5
manipulator.configure(bolt, set_price=5.0)
assert manipulator.inspect(bolt, get_price=None) == 5.0

assert box.to_dict()["items"]["bolt"]["price"] == 5.0

There is no operation layer to write: inspect and configure follow from the request model itself. You write a Super when an operation carries logic of your own:

from msb_arch import Super

class Pricing(Super):
    OPERATION = "price"

    def _price_parts(self, obj, attributes):
        return sum(part.price for part in obj.get_items())

manipulator.register_operation(Pricing(manipulator))
assert manipulator.price(box) == 5.0

A handler is short 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:

answer = manipulator.inspect(bolt, get_price=None, get=["name", "isactive"])
assert answer["get_price"]["result"] == 5.0
assert answer["get"]["result"] == {"name": "bolt", "isactive": True}

Run several requests as one batch:

manipulator.batch([
    {"operation": "configure", "obj": bolt, "attributes": {"set_price": 6.0}},
    {"operation": "inspect", "obj": bolt, "attributes": {"get_price": None}},
])
assert bolt.price == 6.0

Or as a pipeline, when the steps feed each other:

outcome = manipulator.pipeline({
    "written": {"operation": "save", "obj": box, "path": "box.json"},
    "read":    {"operation": "load", "obj": box, "path": "box.json", "after": ["written"]},
    "total":   {"operation": "price", "obj": "@read"},
})
assert outcome.output == 6.0

Architecture

Three layers, plus what they share.

Layer	Module	What lives there
Base — the data	serializable.py, baseentity.py, basecontainer.py	Validation, serialization, caching, ownership
Super — the operations	super.py, builtins.py, project.py	Handlers, method resolution, the built-in operations
Mega — the entry point	manipulator.py	Operation registry, request processing, facades, batches, pipelines
Derivation	catalogue.py, model.py, scaffold.py	What is registered, what holds what, generated stubs
Shared	interceptors.py, results.py, utils/	Metrics and journal, result types, logging, validation

Main classes:

• Serializable — what an entity and a container have in common: annotated fields and their validation, name and isactive, to_dict, the cache, revision and fingerprint.
• BaseEntity — an object addressed by its attributes.
• BaseContainer[T] — a named collection addressed by its items. A sibling of BaseEntity, not a subclass: the two 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 and pipelines, and answers what it knows about itself and the model.
• 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 — requests, pipelines, interceptors, the async surface
• Examples
• Roadmap — what is open, and what was decided against
• 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.

Projects using MSB

• pAstroCORE — radio astronomy observation planning.

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 [email protected] for details.

Contacts

• Author: Alexey Rudnitskiy
• Email: [email protected]
• Repository: https://github.com/Torward1024/MSB
• Version: 1.9.0