mpservice 0.9.9

Serving with asyncio, multiprocessing, and batching

mpservice

Utilities for Python concurrency, including

• Serving with multiprocessing to make full use of multiple cores, and batching to take advantage of vectorized computation if some components of the service have that capability.

  One use case is machine learning model serving, although the code is generic and not restricted to this particular use case.

• Stream processing, i.e. processing a long, possibly infinite stream of input data, with multiple operators in the pipeline. A main use case is that one or more of the operators is I/O bound (think: calling an external service), hence can benefit from concurrency. Both sync and async interfaces are provided.

The serving and streaming utilities can be combined because a mpservice.mpserver.MPServer instance, while doing heavy-lifting in other processes, acts as an I/O bound operator in the main process. Indeed, mpservice.mpserver.MPServer provides methods stream and async_stream for using the server to process data streams.

A MPServer object could be used either in "embedded" mode or to back a HTTP service. In the latter case, the starlette package is a viable choice for providing async service. A few utilities in mpservice.http_server assist with this use case.

The package also contains some other related utilities.

To install, do pip install mpservice.