[python] live output from subprocess command

I'm using a python script as a driver for a hydrodynamics code. When it comes time to run the simulation, I use subprocess.Popen to run the code, collect the output from stdout and stderr into a subprocess.PIPE --- then I can print (and save to a log-file) the output information, and check for any errors. The problem is, I have no idea how the code is progressing. If I run it directly from the command line, it gives me output about what iteration its at, what time, what the next time-step is, etc.

Is there a way to both store the output (for logging and error checking), and also produce a live-streaming output?

The relevant section of my code:

ret_val = subprocess.Popen( run_command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, shell=True )
output, errors = ret_val.communicate()
log_file.write(output)
print output
if( ret_val.returncode ):
    print "RUN failed\n\n%s\n\n" % (errors)
    success = False

if( errors ): log_file.write("\n\n%s\n\n" % errors)

Originally I was piping the run_command through tee so that a copy went directly to the log-file, and the stream still output directly to the terminal -- but that way I can't store any errors (to my knowlege).

My temporary solution so far:

ret_val = subprocess.Popen( run_command, stdout=log_file, stderr=subprocess.PIPE, shell=True )
while not ret_val.poll():
    log_file.flush()

then, in another terminal, run tail -f log.txt (s.t. log_file = 'log.txt').

A good but "heavyweight" solution is to use Twisted - see the bottom.

If you're willing to live with only stdout something along those lines should work:

import subprocess
import sys
popenobj = subprocess.Popen(["ls", "-Rl"], stdout=subprocess.PIPE)
while not popenobj.poll():
   stdoutdata = popenobj.stdout.readline()
   if stdoutdata:
      sys.stdout.write(stdoutdata)
   else:
      break
print "Return code", popenobj.returncode

(If you use read() it tries to read the entire "file" which isn't useful, what we really could use here is something that reads all the data that's in the pipe right now)

One might also try to approach this with threading, e.g.:

import subprocess
import sys
import threading

popenobj = subprocess.Popen("ls", stdout=subprocess.PIPE, shell=True)

def stdoutprocess(o):
   while True:
      stdoutdata = o.stdout.readline()
      if stdoutdata:
         sys.stdout.write(stdoutdata)
      else:
         break

t = threading.Thread(target=stdoutprocess, args=(popenobj,))
t.start()
popenobj.wait()
t.join()
print "Return code", popenobj.returncode

Now we could potentially add stderr as well by having two threads.

Note however the subprocess docs discourage using these files directly and recommends to use communicate() (mostly concerned with deadlocks which I think isn't an issue above) and the solutions are a little klunky so it really seems like the subprocess module isn't quite up to the job (also see: http://www.python.org/dev/peps/pep-3145/ ) and we need to look at something else.

A more involved solution is to use Twisted as shown here: https://twistedmatrix.com/documents/11.1.0/core/howto/process.html

The way you do this with Twisted is to create your process using reactor.spawnprocess() and providing a ProcessProtocol that then processes output asynchronously. The Twisted sample Python code is here: https://twistedmatrix.com/documents/11.1.0/core/howto/listings/process/process.py

None of the Pythonic solutions worked for me. It turned out that proc.stdout.read() or similar may block forever.

Therefore, I use tee like this:

subprocess.run('./my_long_running_binary 2>&1 | tee -a my_log_file.txt && exit ${PIPESTATUS}', shell=True, check=True, executable='/bin/bash')

This solution is convenient if you are already using shell=True.

${PIPESTATUS} captures the success status of the entire command chain (only available in Bash). If I omitted the && exit ${PIPESTATUS}, then this would always return zero since tee never fails.

unbuffer might be necessary for printing each line immediately into the terminal, instead of waiting way too long until the "pipe buffer" gets filled. However, unbuffer swallows the exit status of assert (SIG Abort)...

2>&1 also logs stderror to the file.

Similar to previous answers but the following solution worked for for me on windows using Python3 to provide a common method to print and log in realtime (getting-realtime-output-using-python):

def print_and_log(command, logFile):
    with open(logFile, 'wb') as f:
        command = subprocess.Popen(command, stdout=subprocess.PIPE, shell=True)

        while True:
            output = command.stdout.readline()
            if not output and command.poll() is not None:
                f.close()
                break
            if output:
                f.write(output)
                print(str(output.strip(), 'utf-8'), flush=True)
        return command.poll()

We can also use the default file iterator for reading stdout instead of using iter construct with readline().

import subprocess
import sys
process = subprocess.Popen(your_command, stdout=subprocess.PIPE)
for line in process.stdout:
    sys.stdout.write(line)

Solution 1: Log stdout AND stderr concurrently in realtime

A simple solution which logs both stdout AND stderr concurrently, line-by-line in realtime into a log file.

import subprocess as sp
from concurrent.futures import ThreadPoolExecutor


def log_popen_pipe(p, stdfile):

    with open("mylog.txt", "w") as f:

        while p.poll() is None:
            f.write(stdfile.readline())
            f.flush()

        # Write the rest from the buffer
        f.write(stdfile.read())


with sp.Popen(["ls"], stdout=sp.PIPE, stderr=sp.PIPE, text=True) as p:

    with ThreadPoolExecutor(2) as pool:
        r1 = pool.submit(log_popen_pipe, p, p.stdout)
        r2 = pool.submit(log_popen_pipe, p, p.stderr)
        r1.result()
        r2.result()

Solution 2: A function read_popen_pipes() that allows you to iterate over both pipes (stdout/stderr), concurrently in realtime

import subprocess as sp
from queue import Queue, Empty
from concurrent.futures import ThreadPoolExecutor


def enqueue_output(file, queue):
    for line in iter(file.readline, ''):
        queue.put(line)
    file.close()


def read_popen_pipes(p):

    with ThreadPoolExecutor(2) as pool:
        q_stdout, q_stderr = Queue(), Queue()

        pool.submit(enqueue_output, p.stdout, q_stdout)
        pool.submit(enqueue_output, p.stderr, q_stderr)

        while True:

            if p.poll() is not None and q_stdout.empty() and q_stderr.empty():
                break

            out_line = err_line = ''

            try:
                out_line = q_stdout.get_nowait()
                err_line = q_stderr.get_nowait()
            except Empty:
                pass

            yield (out_line, err_line)

# The function in use:

with sp.Popen(["ls"], stdout=sp.PIPE, stderr=sp.PIPE, text=True) as p:

    for out_line, err_line in read_popen_pipes(p):
        print(out_line, end='')
        print(err_line, end='')

    p.poll()

All of the above solutions I tried failed either to separate stderr and stdout output, (multiple pipes) or blocked forever when the OS pipe buffer was full which happens when the command you are running outputs too fast (there is a warning for this on python poll() manual of subprocess). The only reliable way I found was through select, but this is a posix-only solution:

import subprocess
import sys
import os
import select
# returns command exit status, stdout text, stderr text
# rtoutput: show realtime output while running
def run_script(cmd,rtoutput=0):
    p = subprocess.Popen(cmd, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
    poller = select.poll()
    poller.register(p.stdout, select.POLLIN)
    poller.register(p.stderr, select.POLLIN)

    coutput=''
    cerror=''
    fdhup={}
    fdhup[p.stdout.fileno()]=0
    fdhup[p.stderr.fileno()]=0
    while sum(fdhup.values()) < len(fdhup):
        try:
            r = poller.poll(1)
        except select.error, err:
            if err.args[0] != EINTR:
                raise
            r=[]
        for fd, flags in r:
            if flags & (select.POLLIN | select.POLLPRI):
                c = os.read(fd, 1024)
                if rtoutput:
                    sys.stdout.write(c)
                    sys.stdout.flush()
                if fd == p.stderr.fileno():
                    cerror+=c
                else:
                    coutput+=c
            else:
                fdhup[fd]=1
    return p.poll(), coutput.strip(), cerror.strip()

Based on all the above I suggest a slightly modified version (python3):

• while loop calling readline (The iter solution suggested seemed to block forever for me - Python 3, Windows 7)
• structered so handling of read data does not need to be duplicated after poll returns not-None
• stderr piped into stdout so both output outputs are read
• Added code to get exit value of cmd.

Code:

import subprocess
proc = subprocess.Popen(cmd, shell=True, stdout=subprocess.PIPE,
                        stderr=subprocess.STDOUT, universal_newlines=True)
while True:
    rd = proc.stdout.readline()
    print(rd, end='')  # and whatever you want to do...
    if not rd:  # EOF
        returncode = proc.poll()
        if returncode is not None:
            break
        time.sleep(0.1)  # cmd closed stdout, but not exited yet

# You may want to check on ReturnCode here

import os

def execute(cmd, callback):
    for line in iter(os.popen(cmd).readline, ''): 
            callback(line[:-1])

execute('ls -a', print)

Why not set stdout directly to sys.stdout? And if you need to output to a log as well, then you can simply override the write method of f.

import sys
import subprocess

class SuperFile(open.__class__):

    def write(self, data):
        sys.stdout.write(data)
        super(SuperFile, self).write(data)

f = SuperFile("log.txt","w+")       
process = subprocess.Popen(command, stdout=f, stderr=f)

Executive Summary (or "tl;dr" version): it's easy when there's at most one subprocess.PIPE, otherwise it's hard.

It may be time to explain a bit about how subprocess.Popen does its thing.

(Caveat: this is for Python 2.x, although 3.x is similar; and I'm quite fuzzy on the Windows variant. I understand the POSIX stuff much better.)

The Popen function needs to deal with zero-to-three I/O streams, somewhat simultaneously. These are denoted stdin, stdout, and stderr as usual.

You can provide:

• None, indicating that you don't want to redirect the stream. It will inherit these as usual instead. Note that on POSIX systems, at least, this does not mean it will use Python's sys.stdout, just Python's actual stdout; see demo at end.
• An int value. This is a "raw" file descriptor (in POSIX at least). (Side note: PIPE and STDOUT are actually ints internally, but are "impossible" descriptors, -1 and -2.)
• A stream—really, any object with a fileno method. Popen will find the descriptor for that stream, using stream.fileno(), and then proceed as for an int value.
• subprocess.PIPE, indicating that Python should create a pipe.
• subprocess.STDOUT (for stderr only): tell Python to use the same descriptor as for stdout. This only makes sense if you provided a (non-None) value for stdout, and even then, it is only needed if you set stdout=subprocess.PIPE. (Otherwise you can just provide the same argument you provided for stdout, e.g., Popen(..., stdout=stream, stderr=stream).)

The easiest cases (no pipes)

If you redirect nothing (leave all three as the default None value or supply explicit None), Pipe has it quite easy. It just needs to spin off the subprocess and let it run. Or, if you redirect to a non-PIPE—an int or a stream's fileno()—it's still easy, as the OS does all the work. Python just needs to spin off the subprocess, connecting its stdin, stdout, and/or stderr to the provided file descriptors.

The still-easy case: one pipe

If you redirect only one stream, Pipe still has things pretty easy. Let's pick one stream at a time and watch.

Suppose you want to supply some stdin, but let stdout and stderr go un-redirected, or go to a file descriptor. As the parent process, your Python program simply needs to use write() to send data down the pipe. You can do this yourself, e.g.:

proc = subprocess.Popen(cmd, stdin=subprocess.PIPE)
proc.stdin.write('here, have some data\n') # etc

or you can pass the stdin data to proc.communicate(), which then does the stdin.write shown above. There is no output coming back so communicate() has only one other real job: it also closes the pipe for you. (If you don't call proc.communicate() you must call proc.stdin.close() to close the pipe, so that the subprocess knows there is no more data coming through.)

Suppose you want to capture stdout but leave stdin and stderr alone. Again, it's easy: just call proc.stdout.read() (or equivalent) until there is no more output. Since proc.stdout() is a normal Python I/O stream you can use all the normal constructs on it, like:

for line in proc.stdout:

or, again, you can use proc.communicate(), which simply does the read() for you.

If you want to capture only stderr, it works the same as with stdout.

There's one more trick before things get hard. Suppose you want to capture stdout, and also capture stderr but on the same pipe as stdout:

proc = subprocess.Popen(cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT)

In this case, subprocess "cheats"! Well, it has to do this, so it's not really cheating: it starts the subprocess with both its stdout and its stderr directed into the (single) pipe-descriptor that feeds back to its parent (Python) process. On the parent side, there's again only a single pipe-descriptor for reading the output. All the "stderr" output shows up in proc.stdout, and if you call proc.communicate(), the stderr result (second value in the tuple) will be None, not a string.

The hard cases: two or more pipes

The problems all come about when you want to use at least two pipes. In fact, the subprocess code itself has this bit:

def communicate(self, input=None):
    ...
    # Optimization: If we are only using one pipe, or no pipe at
    # all, using select() or threads is unnecessary.
    if [self.stdin, self.stdout, self.stderr].count(None) >= 2:

But, alas, here we've made at least two, and maybe three, different pipes, so the count(None) returns either 1 or 0. We must do things the hard way.

On Windows, this uses threading.Thread to accumulate results for self.stdout and self.stderr, and has the parent thread deliver self.stdin input data (and then close the pipe).

On POSIX, this uses poll if available, otherwise select, to accumulate output and deliver stdin input. All this runs in the (single) parent process/thread.

Threads or poll/select are needed here to avoid deadlock. Suppose, for instance, that we've redirected all three streams to three separate pipes. Suppose further that there's a small limit on how much data can be stuffed into to a pipe before the writing process is suspended, waiting for the reading process to "clean out" the pipe from the other end. Let's set that small limit to a single byte, just for illustration. (This is in fact how things work, except that the limit is much bigger than one byte.)

If the parent (Python) process tries to write several bytes—say, 'go\n'to proc.stdin, the first byte goes in and then the second causes the Python process to suspend, waiting for the subprocess to read the first byte, emptying the pipe.

Meanwhile, suppose the subprocess decides to print a friendly "Hello! Don't Panic!" greeting. The H goes into its stdout pipe, but the e causes it to suspend, waiting for its parent to read that H, emptying the stdout pipe.

Now we're stuck: the Python process is asleep, waiting to finish saying "go", and the subprocess is also asleep, waiting to finish saying "Hello! Don't Panic!".

The subprocess.Popen code avoids this problem with threading-or-select/poll. When bytes can go over the pipes, they go. When they can't, only a thread (not the whole process) has to sleep—or, in the case of select/poll, the Python process waits simultaneously for "can write" or "data available", writes to the process's stdin only when there is room, and reads its stdout and/or stderr only when data are ready. The proc.communicate() code (actually _communicate where the hairy cases are handled) returns once all stdin data (if any) have been sent and all stdout and/or stderr data have been accumulated.

If you want to read both stdout and stderr on two different pipes (regardless of any stdin redirection), you will need to avoid deadlock too. The deadlock scenario here is different—it occurs when the subprocess writes something long to stderr while you're pulling data from stdout, or vice versa—but it's still there.

The Demo

I promised to demonstrate that, un-redirected, Python subprocesses write to the underlying stdout, not sys.stdout. So, here is some code:

from cStringIO import StringIO
import os
import subprocess
import sys

def show1():
    print 'start show1'
    save = sys.stdout
    sys.stdout = StringIO()
    print 'sys.stdout being buffered'
    proc = subprocess.Popen(['echo', 'hello'])
    proc.wait()
    in_stdout = sys.stdout.getvalue()
    sys.stdout = save
    print 'in buffer:', in_stdout

def show2():
    print 'start show2'
    save = sys.stdout
    sys.stdout = open(os.devnull, 'w')
    print 'after redirect sys.stdout'
    proc = subprocess.Popen(['echo', 'hello'])
    proc.wait()
    sys.stdout = save

show1()
show2()

When run:

$ python out.py
start show1
hello
in buffer: sys.stdout being buffered

start show2
hello

Note that the first routine will fail if you add stdout=sys.stdout, as a StringIO object has no fileno. The second will omit the hello if you add stdout=sys.stdout since sys.stdout has been redirected to os.devnull.

(If you redirect Python's file-descriptor-1, the subprocess will follow that redirection. The open(os.devnull, 'w') call produces a stream whose fileno() is greater than 2.)

Here is a class which I'm using in one of my projects. It redirects output of a subprocess to the log. At first I tried simply overwriting the write-method but that doesn't work as the subprocess will never call it (redirection happens on filedescriptor level). So I'm using my own pipe, similar to how it's done in the subprocess-module. This has the advantage of encapsulating all logging/printing logic in the adapter and you can simply pass instances of the logger to Popen: subprocess.Popen("/path/to/binary", stderr = LogAdapter("foo"))

class LogAdapter(threading.Thread):

    def __init__(self, logname, level = logging.INFO):
        super().__init__()
        self.log = logging.getLogger(logname)
        self.readpipe, self.writepipe = os.pipe()

        logFunctions = {
            logging.DEBUG: self.log.debug,
            logging.INFO: self.log.info,
            logging.WARN: self.log.warn,
            logging.ERROR: self.log.warn,
        }

        try:
            self.logFunction = logFunctions[level]
        except KeyError:
            self.logFunction = self.log.info

    def fileno(self):
        #when fileno is called this indicates the subprocess is about to fork => start thread
        self.start()
        return self.writepipe

    def finished(self):
       """If the write-filedescriptor is not closed this thread will
       prevent the whole program from exiting. You can use this method
       to clean up after the subprocess has terminated."""
       os.close(self.writepipe)

    def run(self):
        inputFile = os.fdopen(self.readpipe)

        while True:
            line = inputFile.readline()

            if len(line) == 0:
                #no new data was added
                break

            self.logFunction(line.strip())

If you don't need logging but simply want to use print() you can obviously remove large portions of the code and keep the class shorter. You could also expand it by an __enter__ and __exit__ method and call finished in __exit__ so that you could easily use it as context.

In addition to all these answer, one simple approach could also be as follows:

process = subprocess.Popen(your_command, stdout=subprocess.PIPE)

while process.stdout.readable():
    line = process.stdout.readline()

    if not line:
        break

    print(line.strip())

Loop through the readable stream as long as it's readable and if it gets an empty result, stop.

The key here is that readline() returns a line (with \n at the end) as long as there's an output and empty if it's really at the end.

Hope this helps someone.

It looks like line-buffered output will work for you, in which case something like the following might suit. (Caveat: it's untested.) This will only give the subprocess's stdout in real time. If you want to have both stderr and stdout in real time, you'll have to do something more complex with select.

proc = subprocess.Popen(run_command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, shell=True)
while proc.poll() is None:
    line = proc.stdout.readline()
    print line
    log_file.write(line + '\n')
# Might still be data on stdout at this point.  Grab any
# remainder.
for line in proc.stdout.read().split('\n'):
    print line
    log_file.write(line + '\n')
# Do whatever you want with proc.stderr here...

I think that the subprocess.communicate method is a bit misleading: it actually fills the stdout and stderr that you specify in the subprocess.Popen.

Yet, reading from the subprocess.PIPE that you can provide to the subprocess.Popen's stdout and stderr parameters will eventually fill up OS pipe buffers and deadlock your app (especially if you've multiple processes/threads that must use subprocess).

My proposed solution is to provide the stdout and stderr with files - and read the files' content instead of reading from the deadlocking PIPE. These files can be tempfile.NamedTemporaryFile() - which can also be accessed for reading while they're being written into by subprocess.communicate.

Below is a sample usage:

        try:
            with ProcessRunner(('python', 'task.py'), env=os.environ.copy(), seconds_to_wait=0.01) as process_runner:
                for out in process_runner:
                    print(out)
        catch ProcessError as e:
            print(e.error_message)
            raise

And this is the source code which is ready to be used with as many comments as I could provide to explain what it does:

If you're using python 2, please make sure to first install the latest version of the subprocess32 package from pypi.

import os
import sys
import threading
import time
import tempfile
import logging

if os.name == 'posix' and sys.version_info[0] < 3:
    # Support python 2
    import subprocess32 as subprocess
else:
    # Get latest and greatest from python 3
    import subprocess

logger = logging.getLogger(__name__)


class ProcessError(Exception):
    """Base exception for errors related to running the process"""


class ProcessTimeout(ProcessError):
    """Error that will be raised when the process execution will exceed a timeout"""


class ProcessRunner(object):
    def __init__(self, args, env=None, timeout=None, bufsize=-1, seconds_to_wait=0.25, **kwargs):
        """
        Constructor facade to subprocess.Popen that receives parameters which are more specifically required for the
        Process Runner. This is a class that should be used as a context manager - and that provides an iterator
        for reading captured output from subprocess.communicate in near realtime.

        Example usage:


        try:
            with ProcessRunner(('python', task_file_path), env=os.environ.copy(), seconds_to_wait=0.01) as process_runner:
                for out in process_runner:
                    print(out)
        catch ProcessError as e:
            print(e.error_message)
            raise

        :param args: same as subprocess.Popen
        :param env: same as subprocess.Popen
        :param timeout: same as subprocess.communicate
        :param bufsize: same as subprocess.Popen
        :param seconds_to_wait: time to wait between each readline from the temporary file
        :param kwargs: same as subprocess.Popen
        """
        self._seconds_to_wait = seconds_to_wait
        self._process_has_timed_out = False
        self._timeout = timeout
        self._process_done = False
        self._std_file_handle = tempfile.NamedTemporaryFile()
        self._process = subprocess.Popen(args, env=env, bufsize=bufsize,
                                         stdout=self._std_file_handle, stderr=self._std_file_handle, **kwargs)
        self._thread = threading.Thread(target=self._run_process)
        self._thread.daemon = True

    def __enter__(self):
        self._thread.start()
        return self

    def __exit__(self, exc_type, exc_val, exc_tb):
        self._thread.join()
        self._std_file_handle.close()

    def __iter__(self):
        # read all output from stdout file that subprocess.communicate fills
        with open(self._std_file_handle.name, 'r') as stdout:
            # while process is alive, keep reading data
            while not self._process_done:
                out = stdout.readline()
                out_without_trailing_whitespaces = out.rstrip()
                if out_without_trailing_whitespaces:
                    # yield stdout data without trailing \n
                    yield out_without_trailing_whitespaces
                else:
                    # if there is nothing to read, then please wait a tiny little bit
                    time.sleep(self._seconds_to_wait)

            # this is a hack: terraform seems to write to buffer after process has finished
            out = stdout.read()
            if out:
                yield out

        if self._process_has_timed_out:
            raise ProcessTimeout('Process has timed out')

        if self._process.returncode != 0:
            raise ProcessError('Process has failed')

    def _run_process(self):
        try:
            # Start gathering information (stdout and stderr) from the opened process
            self._process.communicate(timeout=self._timeout)
            # Graceful termination of the opened process
            self._process.terminate()
        except subprocess.TimeoutExpired:
            self._process_has_timed_out = True
            # Force termination of the opened process
            self._process.kill()

        self._process_done = True

    @property
    def return_code(self):
        return self._process.returncode

If you're able to use third-party libraries, You might be able to use something like sarge (disclosure: I'm its maintainer). This library allows non-blocking access to output streams from subprocesses - it's layered over the subprocess module.