I'm really not sure what this PEP is trying to get at given that it contains no examples and sounds from the descriptions to be adding a complicated api on top of something that already, IMNSHO, has too much it (subprocess.Popen).

Regardless, any user can use the stdout/err/in file objects with their own code that handles them asynchronously (yes that can be painful but that is what is required for _any_ socket or pipe I/O you don't want to block on).

It sounds to me like this entire PEP could be written and released as a third party module on PyPI that offers a subprocess.Popen subclass adding some more convenient non-blocking APIs.  That's where I'd start if I were interested in this as a future feature.


On Fri, Dec 7, 2012 at 5:10 PM, anatoly techtonik <techtonik@gmail.com> wrote:
On Tue, Sep 15, 2009 at 9:24 PM, <exarkun@twistedmatrix.com> wrote:
On 04:25 pm, eric.pruitt@gmail.com wrote:
I'm bumping this PEP again in hopes of getting some feedback.

This is useful, indeed. ActiveState recipe for this has 10 votes, which is high for ActiveState (and such hardcore topic FWIW).
On Tue, Sep 8, 2009 at 23:52, Eric Pruitt <eric.pruitt@gmail.com> wrote:
PEP: 3145
Title: Asynchronous I/O For subprocess.Popen
Author: (James) Eric Pruitt, Charles R. McCreary, Josiah Carlson
Type: Standards Track
Content-Type: text/plain
Created: 04-Aug-2009
Python-Version: 3.2


   In its present form, the subprocess.Popen implementation is prone to
   dead-locking and blocking of the parent Python script while waiting on data
   from the child process.


   A search for "python asynchronous subprocess" will turn up numerous
   accounts of people wanting to execute a child process and communicate with
   it from time to time reading only the data that is available instead of
   blocking to wait for the program to produce data [1] [2] [3].  The current
   behavior of the subprocess module is that when a user sends or receives
   data via the stdin, stderr and stdout file objects, dead locks are common
   and documented [4] [5].  While communicate can be used to alleviate some of
   the buffering issues, it will still cause the parent process to block while
   attempting to read data when none is available to be read from the child


   There is a documented need for asynchronous, non-blocking functionality in
   subprocess.Popen [6] [7] [2] [3].  Inclusion of the code would improve the
   utility of the Python standard library that can be used on Unix based and
   Windows builds of Python.  Practically every I/O object in Python has a
   file-like wrapper of some sort.  Sockets already act as such and for
   strings there is StringIO.  Popen can be made to act like a file by simply
   using the methods attached the the subprocess.Popen.stderr, stdout and
   stdin file-like objects.  But when using the read and write methods of
   those options, you do not have the benefit of asynchronous I/O.  In the
   proposed solution the wrapper wraps the asynchronous methods to mimic a
   file object.

Reference Implementation:

   I have been maintaining a Google Code repository that contains all of my
   changes including tests and documentation [9] as well as blog detailing
   the problems I have come across in the development process [10].

   I have been working on implementing non-blocking asynchronous I/O in the
   subprocess.Popen module as well as a wrapper class for subprocess.Popen
   that makes it so that an executed process can take the place of a file by
   duplicating all of the methods and attributes that file objects have.

"Non-blocking" and "asynchronous" are actually two different things. From the rest of this PEP, I think only a non-blocking API is being introduced.  I haven't looked beyond the PEP, though, so I might be missing something.

I suggest renaming http://www.python.org/dev/peps/pep-3145/ to 'Non-blocking I/O for subprocess' and continue. IMHO on this stage is where examples with deadlocks that occur with current subprocess implementation are badly needed.

   There are two base functions that have been added to the subprocess.Popen
   class: Popen.send and Popen._recv, each with two separate implementations,
   one for Windows and one for Unix based systems.  The Windows
   implementation uses ctypes to access the functions needed to control pipes
   in the kernel 32 DLL in an asynchronous manner.  On Unix based systems,
   the Python interface for file control serves the same purpose.  The
   different implementations of Popen.send and Popen._recv have identical
   arguments to make code that uses these functions work across multiple

Why does the method for non-blocking read from a pipe start with an "_"? This is the convention (widely used) for a private API.  The name also doesn't suggest that this is the non-blocking version of reading. Similarly, the name "send" doesn't suggest that this is the non-blocking version of writing.

The implementation is based on http://code.activestate.com/recipes/440554/ which is more clearly illustrates integrated functionality.

_recv() is a private base function, which is takes stdout or stderr as parameter. Corresponding user-level functions to read from stdout and stderr are .recv() and .recv_err() 

I thought about renaming API to .asyncread() and .asyncwrite(), but that may mean that you call method and then result asynchronously start to fill some buffer, which is not the case here.

Then I thought about .check_read() and .check_write(), literally meaning 'check and read' or 'check and return' for non-blocking calls if there is nothing. But then again, poor naming convention of subprocess uses .check_output() for blocking read until command completes.

Currently, subversion doesn't have .read and .write methods. It may be the best option:
  .write(what)  to pipe more stuff into input buffer of child process.
  .read(from)  where `from` is either subprocess.STDOUT or STDERR
Both functions should be marked as non-blocking in docs and returning None if pipe is closed.

   When calling the Popen._recv function, it requires the pipe name be
   passed as an argument so there exists the Popen.recv function that passes
   selects stdout as the pipe for Popen._recv by default.  Popen.recv_err
   selects stderr as the pipe by default. "Popen.recv" and "Popen.recv_err"
   are much easier to read and understand than "Popen._recv('stdout' ..." and
   "Popen._recv('stderr' ..." respectively.

What about reading from other file descriptors?  subprocess.Popen allows arbitrary file descriptors to be used.  Is there any provision here for reading and writing non-blocking from or to those?

On Windows it is WriteFile/ReadFile and PeekNamedPipe. On Linux it is select. Of course a test is needed, but why it should not just work?
   Since the Popen._recv function does not wait on data to be produced
   before returning a value, it may return empty bytes.  Popen.asyncread
   handles this issue by returning all data read over a given time

Oh.  Popen.asyncread?   What's that?  This is the first time the PEP mentions it.

I guess that's for blocking read with timeout.
Among the most popular questions about Python it is the question number ~500.
   The ProcessIOWrapper class uses the asyncread and asyncwrite functions to
   allow a process to act like a file so that there are no blocking issues
   that can arise from using the stdout and stdin file objects produced from
   a subprocess.Popen call.

What's the ProcessIOWrapper class?  And what's the asyncwrite function? Again, this is the first time it's mentioned.

Oh. That's a wrapper to access subprocess pipes with familiar file API. It is interesting:
So, to sum up, I think my main comment is that the PEP seems to be missing a significant portion of the details of what it's actually proposing.  I suspect that this information is present in the implementation, which I have not looked at, but it probably belongs in the PEP.


Writing PEPs is definitely a job, and a hard one for developers. Too bad a good idea *and* implementation (tests needed) is put on hold, because there is nobody, who can help with that part.

IMHO PEP needs to expand on user stories even if there is significant amount of cited sources, a practical summary and problem illustration by examples are missing.

Python-Dev mailing list
Unsubscribe: http://mail.python.org/mailman/options/python-dev/greg%40krypto.org