My current solution is to setup an IPython ZMQ kernel. I don't just use
IPython.embed_kernel()
because that has many side effects, such as messing around with sys.stdout
, sys.stderr
, sys.excepthook
, signal.signal
, etc and I don't want these side effects. Also, embed_kernel()
is blocking and doesn't really work out-of-the-box in a separate thread (see here).
So, I came up with this code, which is far too complicated in my opinion. (That is why I created a feature request here.)
def initIPythonKernel():
# You can remotely connect to this kernel. See the output on stdout.
try:
import IPython.kernel.zmq.ipkernel
from IPython.kernel.zmq.ipkernel import Kernel
from IPython.kernel.zmq.heartbeat import Heartbeat
from IPython.kernel.zmq.session import Session
from IPython.kernel import write_connection_file
import zmq
from zmq.eventloop import ioloop
from zmq.eventloop.zmqstream import ZMQStream
IPython.kernel.zmq.ipkernel.signal = lambda sig, f: None # Overwrite.
except ImportError, e:
print "IPython import error, cannot start IPython kernel. %s" % e
return
import atexit
import socket
import logging
import threading
# Do in mainthread to avoid history sqlite DB errors at exit.
# https://github.com/ipython/ipython/issues/680
assert isinstance(threading.currentThread(), threading._MainThread)
try:
connection_file = "kernel-%s.json" % os.getpid()
def cleanup_connection_file():
try:
os.remove(connection_file)
except (IOError, OSError):
pass
atexit.register(cleanup_connection_file)
logger = logging.Logger("IPython")
logger.addHandler(logging.NullHandler())
session = Session(username=u'kernel')
context = zmq.Context.instance()
ip = socket.gethostbyname(socket.gethostname())
transport = "tcp"
addr = "%s://%s" % (transport, ip)
shell_socket = context.socket(zmq.ROUTER)
shell_port = shell_socket.bind_to_random_port(addr)
iopub_socket = context.socket(zmq.PUB)
iopub_port = iopub_socket.bind_to_random_port(addr)
control_socket = context.socket(zmq.ROUTER)
control_port = control_socket.bind_to_random_port(addr)
hb_ctx = zmq.Context()
heartbeat = Heartbeat(hb_ctx, (transport, ip, 0))
hb_port = heartbeat.port
heartbeat.start()
shell_stream = ZMQStream(shell_socket)
control_stream = ZMQStream(control_socket)
kernel = Kernel(session=session,
shell_streams=[shell_stream, control_stream],
iopub_socket=iopub_socket,
log=logger)
write_connection_file(connection_file,
shell_port=shell_port, iopub_port=iopub_port, control_port=control_port, hb_port=hb_port,
ip=ip)
print "To connect another client to this IPython kernel, use:", \
"ipython console --existing %s" % connection_file
except Exception, e:
print "Exception while initializing IPython ZMQ kernel. %s" % e
return
def ipython_thread():
kernel.start()
try:
ioloop.IOLoop.instance().start()
except KeyboardInterrupt:
pass
thread = threading.Thread(target=ipython_thread, name="IPython kernel")
thread.daemon = True
thread.start()
Note that this code is outdated now. I have made a package here which should contain a more recent version, and which can be installed via pip.
Other alternatives to attach to running CPython process without having it prepared beforehand. Those usually use the OS debugging capabilities (or use gdb/lldb) to attach to the native CPython process and then inject some code or just analyze the native CPython thread stacks.
Here are other alternatives where you prepare your Python script beforehand to listen on some (tcp/file) socket to provide an interface for remote debugging and/or just a Python shell / REPL.
Some overviews and collected code examples:
(This overview is from here.)