Overhaul logging background thread transport

[theacodes]

No description provided.

[theacodes]

@dhermes ready for another round.

[waprin]

So the excluded loggers change, particularly oath2->google.auth looks very important (maybe the simple explanation for the deadlock). In fact now I'm highly highly highly suspicious that's the root cause of these deadlocks popping up so that should get merged asap.

Just as a tangent, I think the most useful change to make would be to automatically detect recursive logging traces so that type of breakage is impossible. I think Jon gave me some ideas on how that might be possible with logging filters but I haven't dug deep on it yet.

As for switching from the conditions to the queue, I am 99% indifferent but 1% 👎 , I thought about doing it that way but removing all the items from the Queue at once seemed like you might as well just use the Batch entries list, instead of having to basically copy the whole list over every time you commit which seems awkward to me. But obviously you found the existing code ugly and it's mostly a stylistic judgement call so it's up to you and other maintainers, like I said largely indifferent one way or the other since it's more or less accomplishing the same thing.

[theacodes]

As for switching from the conditions to the queue, I am 99% indifferent but 1% 👎 ... But obviously you found the existing code ugly and it's mostly a stylistic judgement call so it's up to you and other maintainers, like I said largely indifferent one way or the other since it's more or less accomplishing the same thing.

My reasoning for doing this wasn't that the code was ugly. I switched to the queue pattern because it's widely deployment, widely tested, well-understood, and easy to reason about. It gets us closer to shared-nothing which is ideal when working with concurrency. I did not like the idea of using locks/conditions/mutexes to patch around a non-thread-safe data structure.

I thought about doing it that way but removing all the items from the Queue at once seemed like you might as well just use the Batch entries list, instead of having to basically copy the whole list over every time you commit which seems awkward to me

We're not draining the entire queue each time, just a subset. And I'm less worried about copying the list because Python essentially just passes around lists of pointers in this case.

[dhermes]

LGTM