🌀 Teleport

Efficiently send large arrays across machines.

Overview

Teleport provides a Server that you can bind functions to and a Client that can call the functions and receive their results. The function arguments and return values are both trees of Numpy arrays. The data is sent efficiently without serialization to maximize throughput.

Installation

pip install git+https://github.com/danijar/teleport.git

Example

This example runs the server and client in the same Python program using subprocesses, but they could also be separate Python scripts running on different machines.

def server():
  import teleport
  server = teleport.Server('tcp://*:2222')
  server.bind('add', lambda data: {'result': data['foo'] + data['bar']})
  server.bind('msg', lambda data: print('Message from client:', data['msg']))
  server.run()

def client():
  import teleport
  client = teleport.Client('tcp://localhost:2222')
  client.connect()
  future = client.add({'foo': 1, 'bar': 1})
  result = future.result()
  print(result)  # {'result': 2}
  client.msg({'msg': 'Hello World'})

if __name__ == '__main__':
  import teleport
  server_proc = teleport.Process(server, start=True)
  client_proc = teleport.Process(client, start=True)
  client_proc.join()
  server_proc.terminate()

Features

Several productivity and performance features are available:

Request batching: The server can batch requests together so that the user function receives a dict of stacked arrays and the function result will be split and sent back to the corresponding clients.
Multithreading: Servers can use a thread pool to process multiple requests in parallel. Optionally, each function can also request its own thread pool to allow functions to block (e.g. for rate limiting) without blocking other functions.
Async clients: Clients can send multiple overlapping requests and wait on the results when needed using Future objects. The maximum number of inflight requests can be limited to avoid requests building up when the server is slower than the client.
Error handling: Exceptions raised in server functions are reported to the client and raised in future.result() or, if the user did not store the future object, on the next request. Worker exception can also be reraised in the server application using server.check().
Heartbeating: Clients can send ping requests when they have not received a result from the server for a while, allowing to wait for results that take a long time to compute without assuming connection loss.
Concurrency: Thread and Process implementations with exception forwarding that can be forcefully terminated by the parent, which Python threads do not natively support. Stoppable threads and processes are also available for coorperative shutdown.
GIL load reduction: The ProcServer behaves just like the normal Server but uses a background process to batch requests and fan out results, substantially reducing GIL load for the server workers in the main process.

Questions

If you have a question, please file an issue.