Debugger crashes with Python 3 project

[PoncinMatthieu]

Summary

We are unable to debug python3 bazel applications when the project sdk is setup to python 3, the debugger will crash with the following error before even executing the application:

Process finished with exit code 137 (interrupted by signal 9: SIGKILL)

If we don't switch the project sdk to python 3, import resolution by intellij will be done using the wrong libs and may be showing unexpected errors.

Steps to reproduce

Clone code example from rules_python repo: https://github.com/bazelbuild/rules_python

Create a venv with python 3, I used 3.6.5 and use the following .bazelproject file:

directories:
  examples/version

targets:
  //examples/version:version_test
  
workspace_type: python

build_flags:
  --python_path=/venv/python3.6.5/bin/python

Debugging will now work if the project sdk is python2, but if you change it to the venv, it will crash.
To change the project sdk, go to File -> Project Structure -> Project

[andrewring]

I'm still seeing this with python 3.6.8.

[mattgodbolt]

This is a super important feature for us in moving to bazel: if there's a sensible set of debugging information that's needed to look in to this, I'm happy to dig a little. I ran the exact command-line IntelliJ runs without the crash; and strace-ing intellij didn't yield any obvious problems.

[mattgodbolt]

Some preliminary investigations, trying to wrap my head around debugging python in intellij in intellij :)

• Seems to trigger when hitting a breakpoint
• Effects seem "racey" - if I single step through the startup code in BlazePyDebugRunner then sometimes it doesn't crash. BUT in those cases breakpoints aren't caught: it's as if the debug stub hasn't waited for the parent before starting running.

I hope this triggers an "aha" moment with someone following this as I'm out of ideas right now!

[danewhite]

On Mac, there seems to be a couple issues at play with the os.execv call within Bazel's python_stub_template.txt.

Digging around on the IntelliJ side of things, I encountered this issue with breakpoints not being triggered by IntelliJ following os.execv (note that my workaround for that still triggers SIGKILL in the Bazel scenario):
https://youtrack.jetbrains.com/issue/PY-37960

I haven't dug through BlazePyDebugRunner at all, but my wild guess would be something involving the timing of the exec call being interpreted as a hangup. Like maybe some socket or file handle getting closed (e.g. FD_CLOEXEC).

A not-so-great workaround is to set IntelliJ to use Python 3.3, which has some different os.execv behavior that doesn't trigger any of these bugs. But if you're using any Python 3.4+ syntax or features, then this won't work for you.

If rules_python decides to implement their proposal for customizing python_stub_template.txt, that would allow a potential workaround by switching os.execv to subprocess.call:
https://github.com/bazelbuild/rules_python/blob/master/proposals/2018-11-08-customizing-the-python-stub-template.md

For an invasive workaround as things currently stand, you can modify your IntelliJ pydev code:
~/Library/Application Support/IntelliJIdea2019.1/python/helpers/pydev/_pydev_bundle/pydev_monkey.py

Change this:

def create_execv(original_name):
    def new_execv(path, args):
        """
        os.execv(path, args)
        os.execvp(file, args)
        """
        import os
        if is_python_args(args):
            send_process_will_be_substituted()
        return getattr(os, original_name)(patch_path(path), patch_args(args))
    return new_execv

To this:

def create_execv(original_name):
    def new_execv(path, args):
        """
        os.execv(path, args)
        os.execvp(file, args)
        """
        from subprocess import call
        call(patch_args(args))
    return new_execv

The remaining downside with this approach for me, is that the debugger opens a new tab with the bazel-out version of the file you're debugging.

[danewhite]

Dug a little more at the Python 3.3 vs 3.4 differences, and there was a change to default all file descriptors to be non-inheritable, including sockets (see the addition of F_DUPFD_CLOEXEC, and https://docs.python.org/3/library/os.html#fd-inheritance). So it's looking like the socket gets closed when os.exec is called.

So a different invasive workaround within IntelliJ's pydev code, is to just set the client socket to be inheritable:
~/Library/Application Support/IntelliJIdea2019.1/python/helpers/pydev/_pydev_bundle/pydev_comm.py

def start_client(host, port):
    """ connects to a host/port """
    pydevd_log(1, "Connecting to ", host, ":", str(port))

    s = socket(AF_INET, SOCK_STREAM)
    # FIX: Set to inheritable to handle changes in Python 3.4
    if hasattr(s, 'set_inheritable'):
        s.set_inheritable(True)

Hopefully IntelliJ will provide a fix of their own sometime soon, but if not, then maybe something on BlazePyDebugRunner that is able to handle the socket being closed and a new one being opened following the os.exec() call.

[danewhite]

So the remaining issue on the Bazel plugin side of things, to open the workspace file instead of the bazel-out file, looks like it's just an issue with symlink resolution.

pydev is passing the symlink path back to the IDE when it triggers the breakpoint. This happens within NetCommandFactory.make_thread_suspend_str() found in the same file as the socket workaround:
~/Library/Application Support/IntelliJIdea2019.1/python/helpers/pydev/_pydev_bundle/pydev_comm.py

Note that the function get_abs_path_real_path_and_base_from_frame returns a tuple of:
(os.abspath(file), os.realpath(file), basename(file))

So just swapping the index used by pydev from 0 to 1 is another invasive workaround.

abs_path_real_path_and_base = get_abs_path_real_path_and_base_from_frame(curr_frame)

# Workaround: Changed from 0 to 1 to pass realpath instead of symlinks
my_file = abs_path_real_path_and_base[1]

if is_real_file(my_file):
    # if filename is Jupyter cell id
    # Workaround: Changed from 0 to 1 to pass realpath instead of symlinks
    my_file = norm_file_to_client(abs_path_real_path_and_base[1])

So with the previous hack of pydev start_client() and this one, the IntelliJ debugger appears to be working correctly.

It seems reasonable to have the IntelliJ team fix the socket issue introduced with Python 3.4, but the symlink issue feels like it probably needs to be handled within the Bazel plugin.

Still haven't dug through this code, but I would hope there's somewhere to resolve the symlink that it receives from pydev.

[danewhite]

Spent a little time looking at the symlink issue, and it looks like there's code for converting the symlink to the workspace file, but that it isn't being used on the file path it receives from the pydev stack frame:

intellij/python/src/com/google/idea/blaze/python/run/BlazePyPositionConverter.java

Lines 50 to 53 in 89e0b25

	@SuppressWarnings("MissingOverride") // added in 2019.1 #api183
	public PySourcePosition convertPythonToFrame(String filePath, int line) {
	return new PySourcePosition(filePath, line) {};
	}

So maybe need to wrap the file path with a call to BlazePyPositionConverter.convertFilePath().

I think this is the IntelliJ upstream call into this method, when parsing the pydev stack frame:
https://github.com/JetBrains/intellij-community/blob/be6247932aa9414ddf7831c0e3becba6940f4839/python/pydevSrc/com/jetbrains/python/debugger/pydev/ProtocolParser.java#L215

[mattgodbolt]

Thanks for your investigations @danewhite. Anyone on the bazel team able to comment if this is the problem, and if so fix? Python2 is about to die and bazel/intellij can't debug python3, which seems a shame (and a blocker to folks like us adopting it across the board)

[sseveran]

Any updates? This has become a problem for us as well.

The fix with set_inheritable worked for us so we have a work around to at least get it to run. The debugger seems to be non-functional after that.

[crutcher-dunnavant-techlabs]

the set_inheritable() bug still appears to be present

[PoncinMatthieu]

The latest bazel version 3.7.0 didn't work with these patches, the bazel generated script is using a wrapper to execute the code called "py3wrapper.sh" and was causing the breakpoints to fail.
For anyone interested, I have made a fork of the pydev debugger to incorporate it all here: https://github.com/PoncinMatthieu/PyDev.Debugger/blob/fix/bazel_debugging/README.md

instructions on the readme:

This branch contains a fix for getting the debugger to work with bazel. This fix works with the following versions:

bazel: 3.7.0
intellij: 2020.1.4
intellij bazel plugin: 2020.10.19.0.3
pydevd: this branch is based on the tag 1.7.1, newer versions didn't work with intellij

Tested only on MacOS.

How to apply the fix?

cd ~/Library/Application\ Support/JetBrains/IntelliJIdea2020.1/plugins/python/helpers/
mv pydev pydev_save
git clone git@github.com:PoncinMatthieu/PyDev.Debugger.git pydev

The fix contains one hard coded variable that you may want to change. Located here. This variable defines the python path used to execute your code.

See this for a diff of the path: fabioz/PyDev.Debugger@main...PoncinMatthieu:fix/bazel_debugging

Do note that this fix is very specific to Bazel so I will not attempt to get this merged on the main pydevd repo.

[sseveran]

@PoncinMatthieu Thanks for such great work. I tested this on Ubuntu 20.04.1 and it worked just fine.

For looking into the future do you have any idea what changes would be needed to the plugin and maybe bazel (specifically rules_python) itself to make a permanent fix? Also I would consider providing some time to making a maintained forked version of the plugin that supports a better python experience. Given the lack of attention these tickets have from Google, they don't internally have the same problems.