Using subprocess.Popen for Process with Large Output

Question

I have some Python code that executes an external app which works fine when the app has a small amount of output, but hangs when there is a lot. My code looks like:

p = subprocess.Popen(cmd, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
errcode = p.wait()
retval = p.stdout.read()
errmess = p.stderr.read()
if errcode:
    log.error('cmd failed <%s>: %s' % (errcode,errmess))

There are comments in the docs that seem to indicate the potential issue. Under wait, there is:

Warning: This will deadlock if the child process generates enough output to a stdout or stderr pipe such that it blocks waiting for the OS pipe buffer to accept more data. Use communicate() to avoid that.

though under communicate, I see:

Note The data read is buffered in memory, so do not use this method if the data size is large or unlimited.

So it is unclear to me that I should use either of these if I have a large amount of data. They don't indicate what method I should use in that case.

I do need the return value from the exec and do parse and use both the stdout and stderr.

So what is an equivalent method in Python to exec an external app that is going to have large output?

Answer 1

A lot of output is subjective so it's a little difficult to make a recommendation. If the amount of output is really large then you likely don't want to grab it all with a single read() call anyway. You may want to try writing the output to a file and then pull the data in incrementally like such:

f=file('data.out','w')
p = subprocess.Popen(cmd, shell=True, stdout=f, stderr=subprocess.PIPE)
errcode = p.wait()
f.close()
if errcode:
    errmess = p.stderr.read()
    log.error('cmd failed <%s>: %s' % (errcode,errmess))
for line in file('data.out'):
    #do something

Answer 2

You're doing blocking reads to two files; the first needs to complete before the second starts. If the application writes a lot to stderr, and nothing to stdout, then your process will sit waiting for data on stdout that isn't coming, while the program you're running sits there waiting for the stuff it wrote to stderr to be read (which it never will be--since you're waiting for stdout).

There are a few ways you can fix this.

The simplest is to not intercept stderr; leave stderr=None. Errors will be output to stderr directly. You can't intercept them and display them as part of your own message. For commandline tools, this is often OK. For other apps, it can be a problem.

Another simple approach is to redirect stderr to stdout, so you only have one incoming file: set stderr=STDOUT. This means you can't distinguish regular output from error output. This may or may not be acceptable, depending on how the application writes output.

The complete and complicated way of handling this is select (http://docs.python.org/library/select.html). This lets you read in a non-blocking way: you get data whenever data appears on either stdout or stderr. I'd only recommend this if it's really necessary. This probably doesn't work in Windows.

Answer 3

You could try communicate and see if that solves your problem. If not, I'd redirect the output to a temporary file.

Answer 4

Glenn Maynard is right in his comment about deadlocks. However, the best way of solving this problem is two create two threads, one for stdout and one for stderr, which read those respective streams until exhausted and do whatever you need with the output.

The suggestion of using temporary files may or may not work for you depending on the size of output etc. and whether you need to process the subprocess' output as it is generated.

As Heikki Toivonen has suggested, you should look at the communicate method. However, this buffers the stdout/stderr of the subprocess in memory and you get those returned from the communicate call - this is not ideal for some scenarios. But the source of the communicate method is worth looking at.

Another example is in a package I maintain, python-gnupg, where the gpg executable is spawned via subprocess to do the heavy lifting, and the Python wrapper spawns threads to read gpg's stdout and stderr and consume them as data is produced by gpg. You may be able to get some ideas by looking at the source there, as well. Data produced by gpg to both stdout and stderr can be quite large, in the general case.