> I'm trying to make sure we track what is blocking people from getting rid of preexec_fn in their existing code so that we can actually deprecate and get rid of the API entirely.

If you consider posix_spawn(), I think that a convenient replacement for preexec_fn function would be a wrapper process which would execute *arbitrary Python code* before spawning the program.

It would not only cover umask case, but also prlimit, and another other custom code.

Pseudo-code of the wrapper:

  import sys
  code = sys.argv[1]
  argv = sys.argv[2:]
  eval(code)
  os.execv(argv[0], argv)

The main risk is that the arbitrary code could create an inheritable file descriptor (not all C extensions respect the PEP 446) which would survive after replacing the process memory with the new program.

Such design would allow to implement it in a third party package (on PyPI) for older Python versions as well.

--

Currently, preexec_fn is a direct reference to a callable Python object in the current process. preexec_fn calls it just after fork().

Here I'm talking about running arbitrary Python code in a freshly spawned Python process. It's different.

--

The new problems are:

* How to ensure that Python is configured as expected? Use -I command line option? Use -S to not import the site module? 
* How to report a Python exception from the child to the parent process? Build a pipe between the two processes and serialize the exception, as we are already doing for preexec_fn?
* How to report os.execv() failure to the parent? Just something like sys.exit(OSErrno.errno)?
* Should close_fds be implemented in the wrapper as well? If yes, can the parent avoid closing file descriptors?

> https://github.com/openstack/oslo.concurrency/blob/master/oslo_concurrency/prlimit.py

This wrapper uses os.execv().