We vendored and modified the callback registry at my day-job (https://github.com/NSLS-II/bluesky/blob/master/bluesky/utils.py#L159) to add a stateful flag to the control if callbacks are allowed to raise or not.

I think that is a better option here as it would require less change to the code base, provides a way to go back to the previous behavior, and does not leave an 'orphaned' method on the callback registry.

The problem of the flag you propose is that exceptions are completely swallowed. An alternative would be to replace the flag by a callback stored as an attribute... say traceback.print_exc for example :-)

Side note for @tacaswell: for your own implementation, given that it is Py3 only you could probably use weakref.WeakMethod instead of having to implement your own _BoundMethodProxy. (And we could do the same once mpl becomes Py3 only...)

@anntzer I agree, that is a better idea than a flag.

👍 to the idea of using a callable attribute

The problem is that we need a half-decent method for locally patching this attribute when needed. However, the only place where this needs to happen is in close_event:

    def close_event(self, guiEvent=None):
        """
        This method will be called by all functions connected to the
        'close_event' with a :class:`CloseEvent`
        """
        s = 'close_event'
        def on_error():
            # Suppress the TypeError when the python session is being killed.
            # It may be that a better solution would be a mechanism to
            # disconnect all callbacks upon shutdown.
            # AttributeError occurs on OSX with qt4agg upon exiting
            # with an open window; 'callbacks' attribute no longer exists.
            tp, value, traceback = sys.exc_info()
            if not isinstance(value, (TypeError, AttributeError)):
                traceback.print_exc()
        event = CloseEvent(s, self, guiEvent=guiEvent)
        self.callbacks.safe_process(s, on_error, event)

(the earlier version wrapped self.callbacks.process to catch these exception specifically). This handling seems to have been introduced by #1098/#1184; I cannot reproduce the issue in #1184, and #1098 is commented as being MacOS only. Can someone with a mac check whether we can get rid of the special silencing there too?

The AttributeError would come out before the process method is called anyway (as it is complaining that callbacks no longer exists). Not sure about the origin of the TypeError though.

Closing in favor of #9063