Unsubscribe all connectivity callbacks on Channel.close

[mehrdada]

Fixes #18995

There could be a better fix available, but seems like the code should be reworked more fundamentally anyway so this simple fix could be fine for now

[mehrdada]

cc @lidizheng

[lidizheng]

Can you try to add a small unit test for this behavior, to prevent regression?

[mehrdada]

@lidizheng it's hard to test against it without refactoring the logic entirely (it'd be easier to fix the entire logic than to test against it)--the reason being the symptom (i.e. the exception) is thrown on a different thread.

[lidizheng]

I agree the text logic probably is more complicate than the fix itself. But it shouldn’t be a challenging one. You can create the callback with the access to the unittest object.

I will label this as help wanted, so if someone have bandwidth can work on it.

[mehrdada]

There's no test for any of these in Channel.close (hence the mess in the first place). If you feel this is not worthwhile progress to fix customer pains short term, then fine. Otherwise, I recommend merging and filing a "help wanted" issue for adding the test.

[lidizheng]

Known failures:

• py27_native.test.fork._fork_interop_test.ForkInteropTest timeout #18305
• 'channelz._channelz_servicer_test.ChannelzServicerTest' is flaky for gevent under Windows #17330