There is a new batch of flakiness attributed to the complexities of shutting down this test. A typical failure is that the execution queue had 1000s of pending requests when the termination sequence starts. There is an apparent corruption about 2 seconds after the test termination is logged, which implies the asynchronous processing of requests is still using the '_callbacks' data structure that should be out of scope by this moment.

The test was initially created to track a very rare race, which later was classified as a race in another test and is already fixed. There is no outstanding issue we expect this test to uncover.