kbus

Status: Alpha · version 0.1.0.

Move messages between parts of an application that may live in the same process, in other processes on the same machine, or on other machines. The calling code is the same in all three cases. Where a part runs is decided at deployment, not in code.

A message is metadata (a small JSON object) plus a payload (bytes). kbus reads the metadata to route the message. It never reads the payload.

Three layers, each built on the one below:

Contents

1. Core: messages, connections, streams
2. Dispatcher: members and routes
3. Link: between applications
4. Failures
5. Limits and backpressure
6. Transports and addresses
7. Test suite
8. What kbus does not do

Core

Message

kbus.Message(meta={"op": "http", "user": "giovanni"}, payload=b"...")

• meta: a JSON object. Bounded in size. Readable by every process the message passes through.
• payload: bytes. Any content, any size up to the frame limit, including empty. kbus copies it, forwards it, never decodes it.

Keys that kbus writes in meta and that an application cannot set: id, kind, route, from, via. Every other key belongs to the application.

Connection

One end of a transport. The other end is another Connection, in the same process or elsewhere.

a, b = kbus.core.pipe()                              # same process
conn = await kbus.core.connect("unix:///tmp/x.sock") # same machine
conn = await kbus.core.connect("wss://host/bus")     # other machine

async for conn in kbus.core.listen("unix:///tmp/x.sock"):
    ...

Each side installs one handler for what arrives:

async def handler(message: kbus.Message, reply: kbus.Reply) -> None:
    ...

conn.handler = handler

Three ways to send

Call — one message, one reply.

reply = await conn.call(message)          # kbus.Message

The handler answers with await reply.send(message) or await reply.fail(exc, meta=None). meta travels with the error reply: the caller finds it, plus kind == "error" and error (the exception class name), in .meta of the exception it gets. A kbus error passed to fail forwards its own .meta by default.

Send — one message, no reply.

await conn.send(message)

await returns when the message is handed to the transport. It waits if the transport's write buffer is full. Delivery is not confirmed. An error in the handler on the other side is logged there and nothing comes back.

Stream — one call that opens a channel. Both sides send sequences of messages on it.

async with conn.open(message) as stream:
    await stream.send(kbus.Message(payload=chunk))   # my direction
    await stream.close()                             # I am done sending
    async for incoming in stream:                    # their direction
        ...

The handler receives the same Stream object as its second argument when the incoming message has kind == "open".

Stream rules:

• Messages arrive in the order they were sent. None is lost.
• A sender waits when the receiver is behind by more than the window (see Limits).
• close() ends my direction. The other direction stays open.
• The stream is finished when both directions are closed.
• abort(reason) ends both directions at once. The other side gets kbus.Aborted(reason) from its next send or iteration.
• Leaving the async with block aborts the stream if it is not finished. This is the deterministic close. Iterating a stream outside async with and stopping early leaves it open until the object is collected.
• If the connection drops, both sides get kbus.LinkLost.

A call is a stream with one message in each direction. The transport carries both the same way.

Many calls at once

A connection carries any number of calls and streams at the same time. Each reply reaches its own caller. Replies may come back in any order.

r1, r2, r3 = await asyncio.gather(conn.call(m1), conn.call(m2), conn.call(m3))

Both directions

Both ends of a connection have a handler and both can call. A handler may call back on the same connection while it is answering.

Timeout and cancellation

async with asyncio.timeout(2):
    reply = await conn.call(message)

On timeout the caller's task is cancelled inside the block and TimeoutError is raised outside it, as the standard library does. On cancellation of any kind kbus sends a cancel to the other side: a handler still running gets asyncio.CancelledError; a stream gets kbus.Aborted. kbus never defines a subclass of CancelledError.

Dispatcher

Members and routes

A Dispatcher holds members. A Member has a name, a secret and a handler. Names are unique while the member is connected.

# the application process
dispatcher = kbus.Dispatcher(secrets={"billing": "b-secret", "shop": "s-secret"})
await dispatcher.listen("unix:///run/app.sock")
await dispatcher.listen("wss://0.0.0.0:8443/bus", ssl=ctx)

# a member, any process
async def handler(message, reply): ...
billing = kbus.Member("billing", secret="b-secret", handler=handler)
await billing.connect("unix:///run/app.sock")   # or connect(dispatcher) in-process

A route is a dotted path. The first segment is a member name. The rest is passed to that member in meta["route"].

reply = await shop.call("billing.total", kbus.Message(payload=b'{"order": 4}'))

The dispatcher reads route, writes from (the verified name of the sender), and forwards the message to billing. It does not decode the payload.

A member that does not exist → kbus.NoSuchMember, immediately.

Identity

from is written by the dispatcher from the name the member proved at connect with its secret. A member cannot set or change it. Any key the application adds to meta is application context, not identity. A policy that decides who may do what reads from.

Nested dispatchers

A member may hold a dispatcher for members below it. A route whose next segment names one of those members is forwarded without touching the payload. Other routes go to the member's own handler.

group = kbus.Member("alfa", secret=..., handler=group_handler,
                    dispatcher=kbus.Dispatcher(secrets=worker_secrets))
await group.connect("wss://commander:8443/bus")
await group.dispatcher.listen("unix:///run/alfa.sock")

• alfa.alfa_0003.http → forwarded to member alfa_0003 of group.dispatcher
• alfa.assign → group_handler

Replies travel back along the same path. meta["via"] lists the members the message passed through.

Hooks

Hooks see metadata only. They cannot change it or the payload.

@dispatcher.on_message
async def policy(meta: dict) -> None:
    if meta["route"].startswith("admin.") and meta["from"] != "root":
        raise kbus.Refused("admin only")       # caller gets kbus.Refused

@group.on_reply
async def observe(meta: dict) -> None:
    ...   # every reply passing through this member, error replies
    ...   # included (meta["kind"] == "error"), before it goes on

@dispatcher.on_disconnect
async def gone(name: str, reason: Exception | None) -> None:
    ...

Objects as members

For members whose handler is a Python object, a convenience on top of raw messages. Only marked methods are reachable. Arguments and results are JSON.

class Billing:
    @kbus.expose
    async def total(self, order: int) -> float: ...

billing = kbus.Member("billing", secret=..., handler=kbus.expose(Billing()))

total = await shop.route("billing").total(order=4)

Unmarked attributes → kbus.NoSuchRoute. Exceptions in the method → kbus.RemoteError with .type, .message, .traceback (text), .meta.

Link

Two applications, each with its own dispatcher, each trusting the other by a token. One connection, opened by one side, carrying traffic in both directions.

# on the laptop, behind NAT
link = kbus.Link("gporcari-mac", token=my_token, url="wss://sourcerer.example/link")
dispatcher.attach_link("sourcerer", link)      # routes "sourcerer:..." go here

# on the server
acceptor = kbus.LinkAcceptor(tokens={my_token: "gporcari-mac", ...})
await acceptor.listen("wss://0.0.0.0:443/link")
dispatcher.attach_acceptor(acceptor)           # each accepted link becomes "<name>:"

Routes across a link carry the instance name before ::

await member.call("sourcerer:kb.ask", message)          # laptop → server
await member.call("gporcari-mac:claude.notify", message) # server → laptop

Rules:

• name:route. A route without : is local.
• Unknown instance name → kbus.NoSuchInstance, immediately.
• One hop only: a route arriving over a link may not name an instance itself; the receiving dispatcher refuses it with kbus.Refused.
• One link per token at a time. A second link presenting the token of a live link is refused and retries with its backoff; it enters once the first link drops.
• The link reconnects by itself after a drop, with backoff. link.reachable says whether it is up now. While it is down, calls to that instance fail immediately with kbus.Unreachable. Nothing is queued.
• On arrival, from is the instance name proved by the token. The members of the other application are not visible as members here.
• What the other application does with a message is decided by its own dispatcher and policies. The link carries application context in meta (for example a user credential); it never carries a resolved identity other than the instance name.

Failures

LinkLost is an uncertain outcome. The message may have been handled before the link dropped. kbus does not retry and does not know. If a call must not run twice, the application makes it idempotent or checks.

A member that dies in the middle of a stream: both sides of the stream get LinkLost; an HTTP response built on it is truncated and the client connection closed.

Limits and backpressure

Every limit is a parameter. They are grouped in one object:

limits = kbus.Limits(
    max_frame=16 * 2**20,      # metadata + payload, bytes
    max_meta=64 * 2**10,       # metadata, bytes
    max_route=4 * 2**10,       # route, bytes
    max_pending=1024,          # calls in flight per connection
    stream_window=32,          # messages in flight per stream
    write_buffer=4 * 2**20,    # transport write buffer, bytes
)

Limits is accepted by Dispatcher, Member, Link, LinkAcceptor, kbus.core.connect and kbus.core.listen. A Member connected in-process without its own Limits uses the dispatcher's. The values above are the defaults.

A limit applies to the side that sets it: a sender does not send what exceeds its own limits, a receiver rejects what exceeds its own. When the two sides differ, the stricter one wins, with an explicit error on the side that detects it.

Link reconnection has its own parameters, on Link:

kbus.Link(..., reconnect=kbus.Reconnect(first=1.0, max=30.0, factor=2.0))

Behaviour at a limit:

• Over max_frame, max_meta or max_route: explicit error, never truncation.
• Over max_pending: Overloaded to the caller, nothing sent.
• Stream window full: stream.send waits for the receiver. Memory per stream is bounded by stream_window × max_frame.
• Write buffer full: send, call and stream.send wait. A slow peer slows its sender; it is never disconnected for being slow.

Data larger than one frame goes through a stream, one message per piece.

An application that embeds kbus exposes these parameters in its own configuration. kajenn does so with an optional kbus: section of the server grammar, one key per field of Limits, and a link: section for Reconnect.

Transports and addresses

Every test in the suite runs on all three. Behaviour is the same; only latency differs.

Test suite

tests/ is the specification. It uses the public API only: no imports of private modules, no inspection of internal state.

• tests/core/ — messages, calls, sends, streams, cancellation, limits, failures. Parametrised on the three transports.
• tests/dispatcher/ — names, secrets, routes, nesting, hooks, from, disconnect of a member with pending calls.
• tests/link/ — token handshake, name: routes, both directions on one connection, drop and reconnect, Unreachable.
• tests/scenarios/ — a minimal HTTP gateway written in the tests, driven by a real HTTP and WebSocket client, with the handler in another process: every HTTP method, bodies of three sizes, a chunked response the client closes halfway, a WebSocket session closed from either side, the handler killed mid-response.

What kbus does not do

• Spawn or supervise processes.
• Discover addresses. Addresses are configuration.
• Queue, persist or retry. What cannot be delivered now fails now.
• Broadcast. Every message has one destination.
• Serialise Python objects. @kbus.expose uses JSON; anything else is bytes.
• Decide trust beyond names: who may call what is a policy you write.