bocpy

Behavior-Oriented Concurrency (BOC) is a new paradigm for parallel and concurrent programming which is particularly well-suited to Python. In a BOC program, data is shared such that each behavior has unique temporal ownership of the data, removing the need for locks to coordinate access. For Python programmers, this brings a lot of benefits. Behaviors are implemented as decorated functions, and from the programmer's perspective, those functions work like normal. Importantly, the programmer's task shifts from solving concurrent data access problems to organizing data flow through functions. The resulting programs are easier to understand, easier to support, easier to extend, and unlock multi-core performance due to the ability to schedule behaviors to run efficiently across multiple sub-interpreters.

BOC has been implemented in several languages, including as a foundational aspect of the research language Verona, and now has been implemented in Python.

Getting Started

You can install bocpy via PyPi:

pip install bocpy

We provide pre-compiled wheels for Python 3.10 onwards on most platforms, but if you have problems with your particular platform/version combination, please file an issue on this repository.

[!NOTE] We provide wheels for Python 3.10 and newer, but bocpy only achieves true parallelism on Python 3.12+, where each sub-interpreter has its own GIL. On 3.10 and 3.11 behaviors still run, but they are serialised by the global GIL. The library may not work on Python versions older than 3.10.

Python version support

The mainline (main) branch in the diagram is the standard CPython build:

• 3.10 / 3.11 — wheels are published and @when works, but every sub-interpreter still shares one process-wide GIL, so behaviors execute one at a time. Use these versions for portability rather than performance.
• 3.12+ — each sub-interpreter gets its own GIL (PEP 684), so worker behaviors run in parallel across cores. This is where bocpy delivers on its concurrency story.
• 3.14 is the current default development and CI target; 3.15 is validated as it stabilises.

The free-threaded branch tracks the no-GIL CPython builds (informally "3.13t", "3.14t", "3.15t" — see PEP 703). bocpy runs unmodified on these interpreters today: we don't re-enable the GIL, and the cown / 2PL protocol gives the same data-race-free guarantees you get on the GIL build. The catch is overhead — on free-threaded Python, the sub-interpreter and XIData machinery is pure ceremony, since plain threads in the main interpreter would already run in parallel.

Issue #5 tracks adding an alternative direct-threading backend that detects a free-threaded interpreter at runtime and skips the sub-interpreter / transpiler / XIData path entirely, while keeping the public Cown / @when API unchanged. We're holding off on that work until the free-threaded build and the relevant CPython APIs stabilise.

Scaling with cores

The chart below shows BOC runtime throughput as the worker count grows from 1 to 8, plotted as speedup relative to a single worker. Numbers come from examples/benchmark.py — a chain-ring workload that exercises the scheduler, two-phase locking, sub-interpreter crossings and the message queue together — run on CPython 3.14 (mean of 3 repeats, 8 s each).

Up to 8 workers, BOC delivers roughly linear scaling on this microbenchmark (≈7.5× at 8 workers). Real applications carry serial costs that this benchmark deliberately strips out — see the docstring at the top of examples/benchmark.py for the load-bearing caveats. To reproduce:

python examples/benchmark.py \
    --sweep-axis workers --sweep-values 1,2,3,4,5,6,7,8 \
    --duration 8 --warmup 2 --repeats 3 \
    --output scaling.json

A behavior can be thought of as a function which depends on zero or more concurrently-owned data objects (which we call cowns). As a programmer, you indicate that you want the function to be called once all of those resources are available. For example, let's say that you had two complex and time-consuming operations, and you needed to act on the basis of both of their outcomes:

def buy_cheese():
    logger = logging.getLogger("cheese_shop")
    for name in all_known_cheeses():
        if is_available(logger, name):
            return name
    
    cleanup_shop(logger)
    return None


def order_meal(exclude: str):
    logger = logging.getLogger("greasy_spoon")
    for dish in menu():
        logger.info(dish)
        if exclude.lower() not in dish.lower():
            logger.info(f"That doesn't have much {exclude} in it")
            return dish

        vikings(logger)
        if random.random() < 0.3:
            logger.info("<bloody vikings>")

    return None


cheese = buy_cheese()
meal = order_meal(exclude="spam")

if meal is not None:
    eat(meal)
elif cheese is not None:
    eat(cheese)

if meal is not None:
    print("I really wanted some cheese...")
elif cheese is not None:
    print("Cheesy comestibles")

return_to_library()

The code above will work, but requires the purveying of cheese and the navigation of the menu for non-spam options to happen sequentially. If we wanted to do these tasks in parallel, we will end up with some version of nested waiting, which can result in deadlock. With BOC, we would write the above like this:

from bocpy import wait, when, Cown

# ...

def buy_cheese():
    cheese = Cown(None)

    @when(cheese)
    def _(cheese):
        logger = logging.getLogger("cheese_shop")
        for name in all_known_cheeses():
            if is_available(logger, name):
                cheese.value = name
                return

        cleanup_shop(logger)

    return cheese


def order_meal(exclude: str):
    order = Cown(None)

    @when(order)
    def _(order, exclude=exclude):
        logger = logging.getLogger("greasy_spoon")
        logger.info("We have...")
        for dish in menu():
            logger.info(dish)
            if exclude.lower() not in dish.lower():
                logger.info(f"That doesn't have much {exclude} in it")
                order.value = dish
                return

            vikings(logger)
            if random.random() < 0.3:
                logger.info("<bloody vikings>")

    return order


cheese = buy_cheese()
meal = order_meal(exclude="spam")


@when(cheese, meal)
def _(cheese, meal):
    if meal.value is not None:
        eat(meal.value)
    elif cheese.value is not None:
        eat(cheese.value)
    else:
        print("<stomach rumbles>")


@when(cheese, meal)
def _(cheese, meal):
    if meal.value is not None:
        print("I really wanted cheese...")
    elif cheese.value is not None:
        print("Cheesy comestibles!")

    return_to_library()


wait()

You can view the full example here

The BOC runtime ensures that this operates without deadlock, by construction.

Talking to main-thread objects

Some values can't survive an XIData round-trip — pyglet shapes, Tk widgets, open file handles, ctypes pointers into a library loaded by __main__, a GPU context. Wrap those in a PinnedCown. Behaviors whose request set contains any pinned cown run on the main thread, drained by pump() from your event loop (or implicitly by wait()).

Keep dispatch coarse — one pinned @when per frame, not per item — so the single-consumer main thread doesn't serialise your worker parallelism. The pump() call drains whatever pinned behaviors are queued and returns immediately when the queue is empty (it never blocks), so it is safe to call from a tight render-loop tick. Hosts that want a starvation warning when the queue stays non-empty can enable it explicitly with set_pump_watchdog(); with no call, the runtime stays silent.

from bocpy import Cown, PinnedCown, pump, start, when

start()
canvas = PinnedCown(MyCanvas())  # main-thread-only handle

def update(dt):
    pump()                                     # drains prior frame's write-back; returns immediately if nothing is queued
    results = [worker_compute(i) for i in range(n)]  # per-item worker @whens

    @when(*results, canvas)                    # one pinned behavior per frame
    def _writeback(*args):
        *cells, canvas = args
        for cell in cells:
            canvas.value.draw(cell.value)

See the Pinned Cowns guide for the coarse-grained dispatch pattern, event-loop integration recipes (pyglet, Tk, asyncio), and the starvation watchdog.

Examples

We provide a few examples to show different ways of using BOC in a program:

1. bocpy-bank: Shows an example where two objects (in this case, bank accounts), interact in an atomic way.
2. bocpy-dining-philosophers: The classic Dining Philosphers problem implemented using BOC.
3. bocpy-fibonacci: A parallel implementation of Fibonacci calculation.
4. bocpy-cooking-boc: The example from the BOC tutorial.
5. bocpy-boids: An agent-based bird flocking example demonstrating the Matrix class for parallel per-cell physics on workers, with one PinnedCown-driven @when per frame batching the pyglet-visible write-back (the coarse-grained pinned-dispatch pattern). Note: you'll need to install pyglet first in order to run the bocpy-boids example.
6. bocpy-primes and bocpy-prime-factor: parallel prime sieve and Pollard's rho factorisation, the latter coordinating early termination via the noticeboard.
7. bocpy-calculator: a small Erlang-style calculator service driven by send/receive.
8. bocpy-cooking-threads: the cooking example written with plain threads, for comparison with bocpy-cooking-boc.
9. bocpy-sketches: the cheese-and-spam sketch shown above as a runnable script.

Why BOC for Python?

Python has always had data races — compound operations like x += 1 are not atomic, even under the GIL — and with the arrival of free-threaded builds (Python 3.13t+) the surface area for concurrency bugs is only growing. BOC eliminates these problems by construction: because behaviors interact with shared data exclusively through cowns, each behavior operates over its data as if it were single-threaded. There is no lock ordering to get right, no forgotten acquire()/release(), and no possibility of deadlock. This holds whether your program runs under the GIL, on per-interpreter GIL (3.12+), or on a free-threaded interpreter.

This library

Our implementation is built on top of the sub-interpreters mechanism and the Cross-Interpreter Data (XIData) API. As of Python 3.12 each sub-interpreter has its own GIL, so behaviors scheduled by bocpy run truly in parallel.

The core scheduling engine is written in C — it is not a wrapper around locks, message queues, or asyncio. Each Cown is backed by a C-level capsule that embeds an MCS-style queue of pending behaviors. When you call @when(a, b), the runtime performs two-phase locking (2PL) over the sorted cown IDs entirely in C (releasing the GIL across the lock-free link loops). Once all cowns in a behavior's request set are acquired, the behavior is dispatched directly to a worker — there is no central scheduler thread and no OS-level lock acquisition on the fast path. Releasing a cown unlinks the MCS node and hands ownership to the next waiting behavior in O(1), which is then dispatched without touching any shared queue. This gives bocpy the same deadlock-freedom-by-construction guarantee as the original Verona runtime.

For cross-behavior data sharing that does not warrant a Cown, the library also provides a small noticeboard — a global key-value store of up to 64 entries. Behaviors can notice_write, notice_update (atomic read-modify-write) and notice_delete keys without acquiring any cowns, and read a frozen snapshot via noticeboard() / notice_read(). The bocpy-prime-factor example uses it to coordinate early termination across worker behaviors.

For values that can't survive an XIData round-trip — UI handles, GPU contexts, file descriptors — the library provides PinnedCown, a cown whose value lives permanently in the main interpreter. Behaviors against a pinned cown run on the main thread, drained by pump() from your event loop or implicitly by wait(). The full surface is PinnedCown, pump, PumpResult, set_pump_watchdog, and set_wait_pump_poll; the bocpy-boids example drives a pyglet window through one pinned @when per frame. See the Pinned Cowns guide for the coarse-grained dispatch pattern, watchdog, and free-threaded support trajectory.

The library also includes lower-level Erlang-style messaging primitives (send / receive) for channel-based communication patterns; see the API documentation for details.

Waiting for completion

Call wait() after scheduling all your behaviors. It blocks the calling thread until every scheduled behavior has finished, then tears down the runtime (joins workers, closes the noticeboard). The next @when call will spin up a fresh runtime automatically.

wait()          # block indefinitely
wait(timeout=5) # raise TimeoutError if not done in 5 s

For a synchronization checkpoint that does not tear the runtime down — e.g. a parallel search that inspects a best-so-far cown between rounds and then continues — use quiesce() instead. It blocks until every in-flight behavior completes, optionally returns a per-worker stats or noticeboard snapshot, and leaves the worker pool and the noticeboard thread running so the next @when call dispatches immediately.

from bocpy import quiesce

snap = quiesce(noticeboard=True)  # dict[str, Any]
print("best so far:", snap.get("best"))
# ... schedule the next round of @when calls ...

Additional Info

BOC is built on a solid foundation of serious scholarship and engineering. For further reading, please see:

1. When Concurrency Matters: Behaviour-Oriented Concurrency
2. Reference implementation in C#
3. OOPSLA23 Talk

C API stability

bocpy is implemented as a CPython C extension that links against the private cross-interpreter data API — _PyXIData_* on 3.14+, _PyCrossInterpreterData_* on 3.12 / 3.13, and on 3.13+ the internal header internal/pycore_crossinterp.h (which requires Py_BUILD_CORE). Under PEP 689 these symbols are explicitly unstable: they may change shape, semantics, or disappear entirely between CPython minor releases, and there is no PyPI / setuptools metadata field that advertises this kind of dependency. The practical consequences are:

• Per-minor wheels. Because we do not target the limited API (Py_LIMITED_API / abi3), every wheel carries a version-specific ABI tag (cp310, cp311, …, cp315). pip will only install a wheel that matches the running interpreter's minor version. The Programming Language :: Python :: 3.x classifiers in pyproject.toml mirror this set.
• Source builds may lag CPython. Alpha / beta / RC builds of a new CPython minor frequently rename or reshape these private symbols. When that happens, bocpy's xidata.h shim needs an update before it will compile against the new headers; until then, install on a released minor version.
• No CPython implementation other than CPython itself. The internal cross-interpreter machinery is CPython-specific, which is why the only implementation classifier we set is Programming Language :: Python :: Implementation :: CPython. PyPy, GraalPy, and other alternatives are not supported.

The compatibility ladder lives in src/bocpy/include/bocpy/xidata.h; the Py_BUILD_CORE #define / #undef save-and-restore there is scoped narrowly to the one #include that needs it, so downstream C extensions that pull in bocpy.h do not inherit it.

Trademarks This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.