subprocess_middleware 0.3

Python 2 and subprocess32

The subprocess module in Python 2.7 can leak file descriptors. Backported fixes from Python 3.x are available in subprocess32, and will be used if installed.

Pipe buffering

For small responses, the unix command cat works as an identity transform. Once a response exceeds the pipe buffer limit (typically 16K or 64K), a deadlock occurs with both processes waiting for the other to read. To avoid this, subprocesses should read in the entirity of each response before writing to stdout and flush stdout at the response end.

Working around this limitation would require writing and reading from different threads.

Apache mod_ext_filter

With mod_ext_filter, the response body is simply piped through an external program. A new process is started for each response, so it suffers from the same limitations as CGI where application setup costs are paid for each request.

FastCGI filter

FastCGI defines a filter role for transforming responses using long-lived processes. Unfortunately the filter role is not supported by Apache mod_fcgid and the FastCGI protocol itself is unnecessarily complicated to implement.

Transforming HTTP reverse proxy

Another alternative would be to implement the transform as part of an HTTP proxy. This adds significant deployment complexity with multiple hops required to support SSL.

PyV8

PyV8 allows running JavaScript in process. It can be tricky to build whereas nodejs packages are easily available.

uWSGI Transformations

uWSGI are working on an rpc plugin for their transformation system. The rpc protocol itself has a 64k request size limit.