C-Two

A resource-oriented RPC framework for Python — turn stateful classes into location-transparent distributed resources.

Basic Idea

Resources, not services — C-Two doesn't expose RPC endpoints. It makes Python classes remotely accessible while preserving their stateful, object-oriented nature.
Zero-copy from process to data — Same-process calls skip serialization entirely. Cross-process IPC can hold shared-memory buffers alive, letting you read columnar data (NumPy, Arrow, …) directly from SHM — no deserialization, no copies.
Built for scientific Python — Native support for Apache Arrow, NumPy arrays, and large payloads (chunked streaming for data beyond 256 MB). Designed for computational workloads, not microservices.
Rust-powered transport — The IPC layer uses a Rust buddy allocator for shared memory and a Rust HTTP relay for high-throughput networking.

Performance

End-to-end cross-process IPC benchmark — same NumPy payload (row_id u32 + x,y,z f64), same machine, same aggregation. Three transport modes compared:

Rows	C-Two hold (ms)	Ray (ms)	C-Two pickle (ms)	Hold vs Ray
1 K	0.07	6.1	0.19	86×
10 K	0.09	7.1	0.82	79×
100 K	0.38	9.8	8.7	26×
1 M	3.7	58	150	15×
3 M	9.7	129	598	13×

C-Two hold — SHM zero-copy via np.frombuffer; no serialization on read
Ray — object store with zero-copy numpy support (Ray 2.55)
C-Two pickle — standard pickle over SHM; included to show serialization cost

Apple M1 Max · Python 3.13 · NumPy 2.4 · See sdk/python/benchmarks/unified_numpy_benchmark.py for full methodology.

Quick Start

pip install c-two

Define a resource contract and its implementation

import c_two as cc

# CRM contract — declares which methods are remotely accessible
@cc.crm(namespace='demo.counter', version='0.1.0')
class Counter:
    def increment(self, amount: int) -> int: ...

    @cc.read
    def value(self) -> int: ...

    def reset(self) -> int: ...


# Resource — a plain Python class implementing the contract
class CounterImpl:
    def __init__(self, initial: int = 0):
        self._value = initial

    def increment(self, amount: int) -> int:
        self._value += amount
        return self._value

    def value(self) -> int:
        return self._value

    def reset(self) -> int:
        old = self._value
        self._value = 0
        return old

Use it locally (zero serialization)

cc.register(Counter, CounterImpl(initial=100), name='counter')
counter = cc.connect(Counter, name='counter')

counter.increment(10)    # → 110
counter.value()          # → 110
counter.reset()          # → 110 (returns old value)
counter.value()          # → 0

cc.close(counter)

Or remotely — same API, different address

# Server process
cc.set_address('ipc:///tmp/my_server')
cc.register(Counter, CounterImpl(), name='counter')

# Client process (separate terminal)
counter = cc.connect(Counter, name='counter', address='ipc:///tmp/my_server')
counter.increment(5)     # works identically
cc.close(counter)

See examples/python/ for complete runnable demos.

Core Concepts

CRM — Contract

A CRM (Core Resource Model) declares which methods a remote resource exposes. It's decorated with @cc.crm(), and method bodies are ... (pure interface — no implementation).

@cc.crm(namespace='demo.greeter', version='0.1.0')
class Greeter:
    @cc.read                              # concurrent reads allowed
    def greet(self, name: str) -> str: ...

    @cc.read
    def language(self) -> str: ...

Methods can be annotated with @cc.read (concurrent access allowed) or left as default write (exclusive access).

Resource — Runtime Instance

A resource is a plain Python class that implements a CRM contract. It holds state and domain logic. It is not decorated — the framework discovers its methods through the CRM contract it was registered under. Name the class by what it is (GreeterImpl, PostgresGreeter, MultilingualGreeter), not the interface.

class GreeterImpl:
    def __init__(self, lang: str = 'en'):
        self._lang = lang
        self._templates = {'en': 'Hello, {}!', 'zh': '你好, {}!'}

    def greet(self, name: str) -> str:
        return self._templates.get(self._lang, 'Hi, {}!').format(name)

    def language(self) -> str:
        return self._lang

Client — Consumer

Anything that calls cc.connect(...) is a client (or consumer / application code). The returned proxy is location-transparent — it works the same whether the resource lives in the same process or on a remote machine.

greeter = cc.connect(Greeter, name='greeter')
greeter.greet('World')     # → 'Hello, World!'
cc.close(greeter)

# Or with context manager:
with cc.connect(Greeter, name='greeter') as greeter:
    greeter.greet('World')

Server — Resource Host

A server is any process that calls cc.register(...) to host one or more resources, then usually cc.serve() to block on the request loop. One server process can host many resources (each under a unique name), and will auto-bind an IPC endpoint the first time a resource is registered.

import c_two as cc

cc.set_address('ipc://my_server')              # optional — default is auto UUID path
cc.register(Greeter, GreeterImpl(), name='greeter')
cc.register(Counter, CounterImpl(), name='counter')
cc.serve()                                     # blocks; Ctrl-C triggers graceful shutdown

Address (ipc://...) is the local transport endpoint. Clients in the same process skip it entirely (zero serialization); clients in a different process on the same host connect to this address directly.
cc.serve() is optional — if your host process has its own event loop (web server, GUI, simulation), you can register resources and let them serve in the background while your main loop runs.
A process can be both a server and a client at the same time (register some resources, connect to others).

Relay — Distributed Discovery

An HTTP relay (c3 relay) is a lightweight broker that lets clients reach servers by name, across machines. Servers announce their IPC address to the relay when they register; clients ask the relay and get routed transparently.

The c3 command is C-Two's cross-language native CLI. From a source checkout, build and link a local development binary with python tools/dev/c3_tool.py --build --link.

# Start a relay anywhere reachable on your network
c3 relay --bind 0.0.0.0:8080

# Server side — announce resources to the relay
cc.set_relay('http://relay-host:8080')
cc.register(MeshStore, MeshStoreImpl(), name='mesh')
cc.serve()

# Client side — resolve by name, no address needed
cc.set_relay('http://relay-host:8080')
mesh = cc.connect(MeshStore, name='mesh')

Multiple relays can form a mesh cluster via gossip — any relay can resolve any resource registered anywhere in the mesh. See the Relay Mesh example below.

When do I need a relay? Only for cross-machine or name-based discovery. Same-process and same-host (IPC) usage work without any relay.

@transferable — Custom Serialization

For custom data types that need to cross the wire, use @cc.transferable. Without it, pickle is used as fallback.

A transferable class defines up to three static methods (written without @staticmethod — the framework adds it automatically):

Method	Required	Purpose
`serialize(data) → bytes`	✅ Yes	Encode data for wire transfer (outbound)
`deserialize(raw) → T`	✅ Yes	Decode wire bytes into an owned Python object (inbound)
`from_buffer(buf) → T`	❌ Optional	Build a zero-copy view over the raw buffer (inbound, hold mode)

import numpy as np

@cc.transferable
class Matrix:
    rows: int
    cols: int
    data: np.ndarray

    def serialize(mat: 'Matrix') -> bytes:
        header = struct.pack('>II', mat.rows, mat.cols)
        return header + mat.data.tobytes()

    def deserialize(raw: bytes) -> 'Matrix':
        rows, cols = struct.unpack_from('>II', raw)
        arr = np.frombuffer(raw, dtype=np.float64, offset=8).reshape(rows, cols)
        return Matrix(rows=rows, cols=cols, data=arr.copy())  # owned copy

    def from_buffer(buf: memoryview) -> 'Matrix':
        header = bytes(buf[:8])
        rows, cols = struct.unpack('>II', header)
        arr = np.frombuffer(buf[8:], dtype=np.float64).reshape(rows, cols)
        return Matrix(rows=rows, cols=cols, data=arr)  # zero-copy view into SHM

When from_buffer is present, the server automatically uses hold mode — the SHM buffer stays alive so from_buffer can return a zero-copy view. Without from_buffer, the server uses view mode — the buffer is released immediately after deserialize.

@cc.transfer — Per-Method Control

Use @cc.transfer() on CRM contract methods to explicitly specify which transferable type handles serialization, or to override the buffer mode:

@cc.crm(namespace='demo.compute', version='0.1.0')
class Compute:
    @cc.transfer(input=Matrix, output=Matrix, buffer='hold')
    def transform(self, mat: Matrix) -> Matrix: ...

    @cc.transfer(input=Matrix, buffer='view')  # force copy even if from_buffer exists
    def ingest(self, mat: Matrix) -> None: ...

Without @cc.transfer, the framework automatically matches registered @transferable types by function signature and resolves the buffer mode from the input type's capabilities.

cc.hold() — Client-Side Zero-Copy

On the client side, cc.hold() requests that the response SHM buffer remain alive, enabling zero-copy reads of the result. The returned HeldResult wraps the value and provides a three-layer safety net for SHM lifecycle:

Explicit .release() — preferred for complex workflows holding multiple buffers
Context manager (with) — recommended for single-buffer scopes
__del__ fallback — last resort, emits ResourceWarning if you forget to release

grid = cc.connect(Compute, name='compute', address='ipc://server')

# Normal call — buffer released immediately after deserialize
result = grid.transform(matrix)

# Option 1: Context manager — clean for single holds
with cc.hold(grid.transform)(matrix) as held:
    data = held.value          # zero-copy NumPy array backed by SHM
    process(data)              # read directly from shared memory
# SHM buffer released on context exit

# Option 2: Explicit release — better for multiple concurrent holds
a = cc.hold(grid.transform)(matrix_a)
b = cc.hold(grid.transform)(matrix_b)
try:
    combined = np.concatenate([a.value.data, b.value.data])
    process(combined)
finally:
    a.release()
    b.release()

When to use hold mode: Large array/columnar data where deserialization dominates cost. For small payloads (< 1 MB), the overhead of tracking SHM lifecycle exceeds the copy cost.

Examples

Single Process — Thread Preference

When cc.connect() targets a CRM registered in the same process, the proxy calls methods directly with zero serialization overhead.

import c_two as cc

cc.register(Greeter, Greeter(lang='en'), name='greeter')
cc.register(Counter, Counter(initial=100), name='counter')

greeter = cc.connect(Greeter, name='greeter')
counter = cc.connect(Counter, name='counter')

print(greeter.greet('World'))    # → Hello, World!
print(counter.value())           # → 100
counter.increment(10)

cc.close(greeter)
cc.close(counter)
cc.shutdown()

Best for: local prototyping, testing, single-machine computation.

Multi-Process IPC — with Custom Transferable

Separate server and client processes communicating over Unix domain sockets with shared memory.

Shared types (types.py):

import c_two as cc
import numpy as np, struct

@cc.transferable
class Mesh:
    n_vertices: int
    positions: np.ndarray   # (N, 3) float64

    def serialize(mesh: 'Mesh') -> bytes:
        header = struct.pack('>I', mesh.n_vertices)
        return header + mesh.positions.tobytes()

    def deserialize(raw: bytes) -> 'Mesh':
        (n,) = struct.unpack_from('>I', raw)
        arr = np.frombuffer(raw, dtype=np.float64, offset=4).reshape(n, 3).copy()
        return Mesh(n_vertices=n, positions=arr)

    def from_buffer(buf: memoryview) -> 'Mesh':
        header = bytes(buf[:4])
        (n,) = struct.unpack('>I', header)
        arr = np.frombuffer(buf[4:], dtype=np.float64).reshape(n, 3)
        return Mesh(n_vertices=n, positions=arr)  # zero-copy view

@cc.crm(namespace='demo.mesh', version='0.1.0')
class MeshStore:
    @cc.read
    def get_mesh(self) -> Mesh: ...

    def update_positions(self, mesh: Mesh) -> int: ...

    @cc.on_shutdown
    def cleanup(self) -> None: ...

Server (server.py):

import c_two as cc
from types import MeshStore, Mesh

class MeshStore:
    def __init__(self):
        self._mesh = Mesh(n_vertices=0, positions=np.empty((0, 3)))

    def get_mesh(self) -> Mesh:
        return self._mesh

    def update_positions(self, mesh: Mesh) -> int:
        self._mesh = mesh
        return mesh.n_vertices

    def cleanup(self):
        print('MeshStore shutting down')

cc.set_address('ipc://mesh_server')
cc.register(MeshStore, MeshStore(), name='mesh')
cc.serve()  # blocks until interrupted

Client (client.py):

import c_two as cc
from types import MeshStore, Mesh
import numpy as np

mesh_store = cc.connect(MeshStore, name='mesh', address='ipc://mesh_server')

# Upload data
big_mesh = Mesh(n_vertices=1_000_000,
                positions=np.random.randn(1_000_000, 3))
mesh_store.update_positions(big_mesh)

# Read with hold — zero-copy SHM access
with cc.hold(mesh_store.get_mesh)() as held:
    positions = held.value.positions  # np.ndarray backed by SHM, no copy
    centroid = positions.mean(axis=0)
    print(f'Centroid: {centroid}')
# SHM released here

cc.close(mesh_store)

Best for: multi-process on same host, worker isolation, high-throughput local IPC.

Cross-Machine — HTTP Relay

An HTTP relay bridges network requests to CRM processes running on IPC. CRM processes register with the relay, and clients discover resources by name.

CRM Server (resource.py):

import c_two as cc

cc.set_relay('http://relay-host:8080')
cc.set_address('ipc://mesh_server')
cc.register(MeshStore, MeshStore(), name='mesh')
cc.serve()  # blocks until Ctrl-C

Relay — start via the c3 CLI:

# Bind address from CLI flag, env var C2_RELAY_BIND, or .env file
c3 relay --bind 0.0.0.0:8080

Client (client.py):

import c_two as cc

cc.set_relay('http://relay-host:8080')
mesh = cc.connect(MeshStore, name='mesh')  # relay resolves the name
mesh.get_mesh()
cc.close(mesh)

Best for: network-accessible services, web integration, cross-machine deployment.

Relay Mesh — Multi-Relay Clusters

Multiple relays form a mesh network with gossip-based route propagation. CRMs register with their local relay; clients discover resources across the entire cluster.

# Start relay A (seeds point to peer relays for auto-join)
c3 relay --bind 0.0.0.0:8080 --relay-id relay-a \
    --advertise-url http://relay-a:8080 --seeds http://relay-b:8080

# Start relay B
c3 relay --bind 0.0.0.0:8080 --relay-id relay-b \
    --advertise-url http://relay-b:8080 --seeds http://relay-a:8080

CRM processes register with their local relay; the mesh propagates routes automatically. Clients can connect through any relay in the mesh.

Best for: multi-node clusters, high availability, geographic distribution.

See examples/python/relay_mesh/ for a complete runnable mesh demo.

Server-Side Monitoring

Use cc.hold_stats() to monitor SHM buffers held by resource methods in hold mode:

stats = cc.hold_stats()
# {'active_holds': 3, 'total_held_bytes': 52428800, 'oldest_hold_seconds': 12.5}

Architecture

The design philosophy of C-Two is not to define services, but to empower resources.

In scientific computation, resources encapsulating complex state and domain-specific operations need to be organized into cohesive units. We call the contracts describing these resources Core Resource Models (CRMs). Applications care more about how to interact with resources than where they are located. C-Two provides location transparency and uniform resource access, so any client can interact with a resource as if it were a local object.

Client Layer

Any code that calls cc.connect(...) to consume a resource. The returned proxy provides full type safety and location transparency — clients don't know (or care) where the resource is running.

cc.connect(CRMClass, name='...', address='...') returns a typed CRM proxy
The proxy supports context management: with cc.connect(...) as x: auto-closes

Resource Layer

Server-side stateful instances exposed through standardized CRM contracts.

CRM contract: Interface class decorated with @cc.crm(). Only methods declared here are remotely accessible.
Resource: Plain Python class implementing the contract — state + domain logic. Not decorated.
@transferable: Custom serialization for domain data types. Optionally provides from_buffer for zero-copy SHM views.
@cc.transfer: Per-method control over input/output transferable types and buffer mode.
@cc.read / @cc.write: Concurrency annotations — parallel reads, exclusive writes.
@cc.on_shutdown: Lifecycle callback invoked when a resource is unregistered (not exposed via RPC).

Transport Layer

Protocol-agnostic communication with automatic protocol detection based on address scheme:

Scheme	Transport	Use case
`thread://`	In-process direct call	Zero serialization, testing
`ipc:///path`	Unix domain socket + shared memory	Multi-process, same host
`http://host:port`	HTTP relay	Cross-machine, web-compatible

The IPC transport uses a control-plane / data-plane separation: method routing flows through UDS inline frames while payload bytes are exchanged via shared memory — zero-copy on the data path. When from_buffer is available, hold mode keeps the SHM buffer alive across the CRM method call, enabling the CRM to operate directly on shared memory without deserialization.

Rust Native Layer

Performance-critical components are implemented in Rust and compiled as a Python extension via PyO3 + maturin:

The Rust workspace contains 7 crates organized in 4 layers (foundation → protocol → transport → bridge):

Buddy Allocator — Zero-syscall shared memory allocation for the IPC transport. Cross-process, lock-free on the fast path.
Wire Protocol — Frame encoding, chunk assembly, and chunk registry for large-payload lifecycle management.
HTTP Relay — High-throughput axum-based gateway bridging HTTP to IPC. Handles connection pooling and request multiplexing.

The Rust extension is compiled automatically during pip install c-two (from pre-built wheels) or uv sync (from source).

The c3 command is distributed as a native CLI binary and built from the root cli/ package. Source checkouts can link a local development binary with python tools/dev/c3_tool.py --build --link; published CLI artifacts are owned by the CLI release pipeline rather than by any language SDK.

Installation

From PyPI

pip install c-two

Pre-built wheels are available for:

Linux: x86_64, aarch64
macOS: Apple Silicon (aarch64), Intel (x86_64)
Python: 3.10, 3.11, 3.12, 3.13, 3.14, 3.14t (free-threading)

If no pre-built wheel is available for your platform, pip will build from source (requires a Rust toolchain).

Development Setup

git clone https://github.com/world-in-progress/c-two.git
cd c-two
cp .env.example .env               # configure environment (optional)
uv sync                            # install dependencies + compile Rust extensions
uv sync --group examples           # install examples dependencies (pandas, pyarrow)
cargo build --manifest-path cli/Cargo.toml  # build the native c3 CLI in source checkouts
uv run pytest                      # run the test suite

Requires uv and a Rust toolchain.

Roadmap

Feature	Status
Core RPC framework (CRM + Resource + Client)	✅ Stable
IPC transport with SHM buddy allocator	✅ Stable
HTTP relay (Rust-powered)	✅ Stable
Relay mesh with gossip-based discovery	✅ Stable
Chunked streaming (payloads > 256 MB)	✅ Stable
Heartbeat & connection management	✅ Stable
Read/write concurrency control	✅ Stable
Unified config architecture (Python SSOT)	✅ Stable
CI/CD & multi-platform PyPI publishing	✅ Stable
Disk spill for extreme payloads	✅ Stable
Hold mode with `from_buffer` zero-copy	✅ Stable
SHM residence monitoring (`cc.hold_stats()`)	✅ Stable
Async interfaces	🔜 Planned
Cross-language clients (TypeScript/Rust)	🔮 Future

See the full roadmap for details.

License

MIT

Built for scientific Python. Powered by Rust.

Name		Name	Last commit message	Last commit date
Latest commit History 684 Commits
.github		.github
.vscode		.vscode
cli		cli
core		core
docs		docs
examples		examples
sdk/python		sdk/python
tools/dev		tools/dev
.env.example		.env.example
.gitignore		.gitignore
.python-version		.python-version
CHANGELOG.md		CHANGELOG.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
README.md		README.md
README.zh-CN.md		README.zh-CN.md
pyproject.toml		pyproject.toml
uv.lock		uv.lock

Folders and files

Latest commit

History

Repository files navigation

C-Two

Basic Idea

Performance

Quick Start

Define a resource contract and its implementation

Use it locally (zero serialization)

Or remotely — same API, different address

Core Concepts

CRM — Contract

Resource — Runtime Instance

Client — Consumer

Server — Resource Host

Relay — Distributed Discovery

@transferable — Custom Serialization

@cc.transfer — Per-Method Control

cc.hold() — Client-Side Zero-Copy

Examples

Single Process — Thread Preference

Multi-Process IPC — with Custom Transferable

Cross-Machine — HTTP Relay

Relay Mesh — Multi-Relay Clusters

Server-Side Monitoring

Architecture

Client Layer

Resource Layer

Transport Layer

Rust Native Layer

Installation

From PyPI

Development Setup

Roadmap

License

About

Topics

Resources

License

Contributing

Uh oh!

Stars

Watchers

Forks

Releases

Uh oh!

Contributors

Uh oh!

Languages