`container.attach(stream=True)` causing "ResourceWarning: unclosed <socket.socket …>"

[Dadeos-Menlo]

Attempting to use container.attach(stream=True) to stream logs from a container appears to leak unclosed sockets.

The following test case:

import contextlib
import unittest

import docker


class TestDocker(unittest.TestCase):

   def test(self):
      for count in range(10):
         with contextlib.closing(docker.from_env()) as client:
            container = client.containers.run('alpine',
               auto_remove=True,
               command=('/bin/sh', '-c', 'echo Hello; sleep 1'),
               detach=True,
               init=True,
               tty=True
            )
            with contextlib.closing(container.attach(
               logs=True,
               stdout=True,
               stream=True
            )) as logs:
               next(logs, b'').decode()
            container.stop()

yields the following ouput:

test (container.test_docker.TestDocker) ... /usr/lib/python3.8/email/feedparser.py:158: ResourceWarning: unclosed <socket.socket [closed] fd=7, family=AddressFamily.AF_UNIX, type=SocketKind.SOCK_STREAM, proto=0>
  _factory(policy=self.policy)
ResourceWarning: Enable tracemalloc to get the object allocation traceback
/usr/lib/python3.8/email/feedparser.py:158: ResourceWarning: unclosed <socket.socket [closed] fd=8, family=AddressFamily.AF_UNIX, type=SocketKind.SOCK_STREAM, proto=0>
  _factory(policy=self.policy)
ResourceWarning: Enable tracemalloc to get the object allocation traceback
/usr/lib/python3.8/email/feedparser.py:158: ResourceWarning: unclosed <socket.socket [closed] fd=9, family=AddressFamily.AF_UNIX, type=SocketKind.SOCK_STREAM, proto=0>
  _factory(policy=self.policy)
ResourceWarning: Enable tracemalloc to get the object allocation traceback
/usr/lib/python3.8/email/feedparser.py:158: ResourceWarning: unclosed <socket.socket [closed] fd=10, family=AddressFamily.AF_UNIX, type=SocketKind.SOCK_STREAM, proto=0>
  _factory(policy=self.policy)
ResourceWarning: Enable tracemalloc to get the object allocation traceback
ok

----------------------------------------------------------------------
Ran 1 test in 4.980s

OK

The root cause appears to relate to a combination of APIClient._read_from_socket(…) and CancellableStream:

docker-py/docker/api/container.py

Lines 61 to 65 in a365202

	output = self._read_from_socket(
	response, stream, self._check_is_tty(container), demux=demux)
	if stream:
	return CancellableStream(output, response)

The documentation for APIClient._read_from_socket(…) states:

If stream=True, then a generator is returned instead and the caller is responsible for closing the response.

and if stream is not True then the implementation calls response.close():

docker-py/docker/api/client.py

Lines 443 to 447 in a365202

	try:
	# Wait for all frames, concatenate them, and return the result
	return consume_socket_output(gen, demux=demux)
	finally:
	response.close()

however, the current implementation of CancellableStream.close() does not call self._response.close().

Modifying the test case to be:

import contextlib
import unittest

import docker


class TestDocker(unittest.TestCase):

   def test(self):
      for count in range(10):
         with contextlib.closing(docker.from_env()) as client:
            container = client.containers.run('alpine',
               auto_remove=True,
               command=('/bin/sh', '-c', 'echo Hello; sleep 1'),
               detach=True,
               init=True,
               tty=True
            )
            logs = container.attach(
               logs=True,
               stdout=True,
               stream=True
            )
            next(logs, b'').decode()
            logs._response.close()
            container.stop()

(i.e. not bothering to call CancellableStream.close(), but manually calling CancellableStream._response.close() appears to resolve the "ResourceWarning: unclosed <socket.socket …>" warnings; which appears to suggest that CancellableStream.close() should call self._response.close() and perhaps the rest of the current implementation that appears to be seeking a socket in order to close it is unnecessary?

This issue may, or may not, be that same as that reported in #3268.

[villaflaminio]

I am experiencing the same problem using version docker==7.1.0 .
My use case is slightly different but we can simplify it as:

logs = container.logs(stdout=True, stderr=True, stream=True)
count = 0
for line in logs:
    print(line.strip())
    count+=1
    if count >= 10:
        break

container.stop()
container.remove()

if I add the logs._response.close(), I no longer encounter the problem