Support debugging of indirect child processes.

Inizia esaminando gli entry point menzionati debugpy.server.cli.options, debugpy.server.api e debugpy.connect, quindi confrontali con il supporto multiprocesso esistente per i processi figli diretti. Definisci il confine ufficiale dell’API per trasferire il contesto della sessione attraverso processi indiretti non Python; il lavoro è completato quando l’albero di worker multilivello mostrato può connettersi senza dipendere dai dettagli di implementazione.

While debugpy natively supports multiprocess debugging, this works only for directly launched child processes.
Sometimes, though, there may be several levels on non-Python processes between the main process and the child one needs to debug. I am hitting this limitation while trying to debug a user-defined function, that runs in the context of a PySpark worker process. The process tree looks like this:

<python process hosting my main program>
 └ <pyspark process>
    └ <yet another pyspark process>
       ├ <python worker process hosting the UDFs>
       ├ <python worker process hosting the UDFs>
       ├ ...
       └ <python worker process hosting the UDFs>

After spelunking the innards of debugpy for a few hours, I've managed to come up with this code:

import pickle
import os
import debugpy

def capture_session_context():
    """To be called in the main process to capture context of the current debugpy session."""
    options = debugpy.server.cli.options
    context = pickle.dumps((options.address, options.adapter_access_token, os.getpid()), 0)
    os.environ["DEBUGPY_CONTEXT"] = context.decode('ascii')

def connect_to_session():
    """To be called in a worker process to connect to the parent session."""
    context = os.environ['DEBUGPY_CONTEXT'].encode('ascii')
    address, token, ppid = pickle.loads(context)
    from debugpy.server import api
    api.pydevd.SetupHolder.setup = {"ppid": ppid}
    debugpy.connect(address, access_token=token)

Obviously, this hack uses a bunch of implementation details. It would be nice if this scenario was officially supported.
(I mean exposing official APIs similar to the above, not that debugpy should support PySperk specifically.