ezq

ezq

Simple wrappers for python multiprocessing and threading.

Changelog - Issues - Documentation

Why?

ezq makes it easy to connect subprocesses and threads (both considered "workers") using queues with a simpler API than concurrent.futures, multiprocessing, or threading.

Install

python -m pip install ezq

Example: Quick Start

If you just want to apply a function to some inputs, you can use ezq.map() to run it on all available CPUs and get the results back.

import ezq
print(list(ezq.map(lambda x: x * 2, range(6))))
# => [0, 2, 4, 6, 8, 10]

Example: Sum Messages

Here's a simple example of a worker that reads from an input queue, sums up the messages, and puts the result on an output queue.

import ezq


def worker(q, out):
    """Add up all the messages."""
    total = 0
    for msg in q:  # read a message from the queue
        total += msg.data

    # after reading all the messages, write the total
    out.put(total)


def main():
    """Run several workers."""
    # Step 1: Creates the queues and start the workers.
    q, out = ezq.Q(), ezq.Q()  # input & output queues
    workers = [ezq.run(worker, q, out) for _ in range(ezq.NUM_CPUS)]
    # workers are all running

    # Step 2: Send work to the workers.
    for i in range(1000):
        q.put(i)  # send work

    # Step 3: Tell the workers to finish.
    q.stop(workers)
    # workers are all stopped

    # Step 4: Process the results.
    want = sum(range(1000))
    have = sum(msg.data for msg in out.items())
    assert have == want
    print(have)


if __name__ == "__main__":
    main()

Typical worker lifecycle

• The main process creates queues with ezq.Q.

• The main process creates workers with ezq.run (alias for Worker.process) or ezq.run_thread (alias for Worker.thread).

• The main process sends data using Q.put.

• The worker iterates over the queue.

• The main process ends the queue with Q.stop.

• The worker returns when it reaches the end of the queue.

• (Optional) The main process processes the results.

Process vs Thread

ezq supports two kinds of workers: Process and Thread. There is a lot of existing discussion about when to use which approach, but a general rule of thumb is:

• Process is for parallelism so you can use multiple CPUs at once. Ideal for CPU-bound tasks like doing lots of mathematical calculations.

• Thread is for concurrency so you can use a single CPU to do multiple things. Ideal for I/O-bound tasks like waiting for a disk, database, or network.

Some more differences:

• Shared memory: Each Process worker has data sent to it via pickle (actually dill, a pickle replacement) and it doesn't share data with other workers. By contrast, each Thread worker shares its memory with all other workers on the same CPU, so it can accidentally change global state.

• Queue overhead: ezq.Q has more overhead for Process workers than Thread workers.

• Creating sub-workers: Process and Thread workers can create additional Thread workers, but they cannot create additional Process workers.

Create queues

In the main process, create the queues you'll need. Here are my common situations:

• 0 queues: I'm using a simple function and can ask ezq.map to make the queues for me.

• 1 queue: the worker reads from an input queue and persists the result somewhere else (e.g., writing to disk, making a network call, running some other program).

• 2 queues (most common): the worker reads from an input queue and write the results to an output queue.

• 3 queues: multiple stages of work are happening where workers are reading from one queue and writing to another queue for another worker to process.

NOTE: If you're using Thread workers, you can save some overhead by passing Q("thread"). This lightweight queue also doesn't use pickle, so you can use it to pass hard-to-pickle things (e.g., database connection).

q, out = ezq.Q(), ezq.Q() # most common
q2 = ez.Q("thread") # only ok for Thread workers

A worker task is just a function

In general, there's nothing special about a worker function, but note:

• If you're using Process workers, all arguments are passed through pickle first.

• We don't currently do anything with the return value of this function (unless you use ezq.map()). You'll need an output queue to return data back to the main process/thread.

Create workers

In the main process, create workers using ezq.run or ezq.run_thread which take a function and any additional parameters. Typically, you'll pass the queues you created to the workers at this point.

NOTE: Process and Thread workers can create additional Thread workers, but they cannot create additional Process workers.

Send data

Once you've created the workers, you send them data with Q.put which creates ezq.Msg objects and puts them in the queue. Each message has three attributes (all optional):

• data: Any - This is the data you want the worker to work on.

• kind: str - You can use this to send multiple kinds of work to the same worker. Note that the special END kind is used to indicate the end of a queue.

• order: int - This is the message order which can help you reorder results or ensure that messages from a queue are read in a particular order (that's what Q.sorted() uses).

Beware pickle

If you are using Process workers, everything passed to the worker (arguments, messages) is first passed to pickle (actually, dill). Anything that cannot be pickled with dill (e.g., database connections), cannot be passed to Process workers. Note that dill can serialize many more types than pickle (e.g. lambda functions).

Beware shared state

If you are using Thread workers, workers can share certain variables, so you need to be careful of how variables are access to avoid accidentally corrupting data.

Iterate over messages

Inside the worker, iterate over the queue to read each message until the queue ends (see below). If the messages need to be processed in order, use Q.sorted.

for msg in q: # read each message until the queue ends
  ...

for msg in q.sorted(): # read each message in order
  ...

End the queue

After the main process has sent all the data to the workers, it needs to indicate that there's no additional work to be done. This is done by calling Q.stop() using the input queue that the workers are reading from and passing the list of workers to wait for.

In some rare situations, you can use Q.end() to explicitly end the queue.

Process results

If you have an output queue, you may want to to process the results. You can use Q.items() to end the queue and read the current messages.

import ezq
out = ezq.Q()
...
result = [msg.data for msg in out.items()]
# OR
result = [msg.data for msg in out.items(sort=True)] # sorted by Msg.order
# OR
result = [msg.data for msg in out.items(cache=True)] # cache the messages

Example: Read and Write Queues

In this example, several workers read from a queue, process data, and then write to a different queue that a single worker uses to print to the screen sorting the results as it goes along.

Note that we use a single writer to avoid clashes or overwriting.

import ezq


def printer(out: ezq.Q) -> None:
    """Print results in increasing order."""
    for msg in out.sorted():
        print(msg.data)


def collatz(q: ezq.Q, out: ezq.Q) -> None:
    """Read numbers and compute values."""
    for msg in q:
        num = float(msg.data)
        if msg.kind == "EVEN":
            out.put((num, num / 2), order=msg.order)
        elif msg.kind == "ODD":
            out.put((num, 3 * num + 1), order=msg.order)


def main() -> None:
    """Run several threads with a subprocess for printing."""
    q, out = ezq.Q("thread"), ezq.Q()
    readers = [ezq.run_thread(collatz, q, out) for _ in range(ezq.NUM_THREADS)]
    writer = ezq.run(printer, out)

    for num in range(40):
        kind = "EVEN" if num % 2 == 0 else "ODD"
        q.put(num, kind=kind, order=num)

    q.stop(readers)
    out.stop(writer)


if __name__ == "__main__":
    main()

License

MIT License