Brian Hulette

It looks like codecov is unhappy since this file looks untested. Presumably because we don't have pytorch installed in the run generating coverage data. I confirmed this test is exercised...

IIUC the appropriate fix is to adjust client libraries used in various IOs to use a shared channel (@lukecwik can advise).

@reuvenlax maybe you're thinking of https://github.com/apache/beam/pull/16619 which fixed a different thread explosion in `RetryManager`?

> Is there any documentation on where we use GAX, or a pointer to a design pattern for this? I'm not sure about documentation. My understanding from offline discussion with...

Is P1 appropriate here? This and other playground tasks look like normal feature work, which should be categorized as P2.

R: @tvalentyn CC: @robertwb

@ryanthompson591 I can't comment on your doc so I'll mention here - it encourages users to materialize the iterable result to a list, which can lead to OOMs. I think...

Run Python PreCommit

`org.apache.beam.sdk.transforms.ParDoLifecycleTest.testTeardownCalledAfterExceptionInStartBundleStateful` is tracked in #21266 Maybe this should just track the GroupByKey one? It does look quite flaky: 

Updated the description so this just tracks the GroupByKeyTest failure. Let's use https://github.com/apache/beam/issues/21266 to track the ParDoLifecycleTest flake