Skip to content

danijar/portal

Repository files navigation

PyPI

🌀 Portal

Fast and reliable distributed systems in Python.

Features

  • 📡 Communication: Portal lets you bind functions to a Server and call them from one or more Clients. Wait on results via Future objects. Clients can automatically restore broken connections.
  • 🚀 Performance: Optimized for throughput and latency. Array data is zero-copy serialized and deserialized for throughput near the hardware limit.
  • 🤸 Flexibility: Function inputs and outputs can be nested dicts and lists of numbers, strings, bytes, None values, and Numpy arrays. Bytes allow applications to chose their own serialization, such as pickle.
  • 🚨 Error handlings: Provides Process and Thread objects that can reliably be killed by the parent. Unhandled exceptions in threads stop the program. Error files can be used to stop distributed systems.
  • 📦 Request batching: Use BatchServer to collect multiple incoming requests and process them at once, for example for AI inference servers. Batching and dispatching happens in a separate process to free the GIL.
  • Correctness: Covered by over 100 unit tests for common usage and edge cases and used for large scale distributed AI systems.

Installation

pip install portal

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 portal
  server = portal.Server(2222)
  server.bind('add', lambda x, y: x + y)
  server.bind('greet', lambda msg: print('Message from client:', msg))
  server.start()

def client():
  import portal
  client = portal.Client('localhost', 2222)
  future = client.add(12, 42)
  result = future.result()
  print(result)  # 54
  client.greet('Hello World')

if __name__ == '__main__':
  import portal
  server_proc = portal.Process(server, start=True)
  client_proc = portal.Process(client, start=True)
  client_proc.join()
  server_proc.kill()
  print('Done')

Questions

Please open a separate GitHub issue for each question.