These tests (especially unconstrained versions) can get very
   backlogged and may take a while to finish. We sometimes flake waiting
   for that. This is not hazardous (IMO), as the scripts that run these
   tests already have timeouts to make sure that these don't truly go on
   forever.

2. Make the time spent in the benchmark phase actually be
   benchmark_seconds rather than benchmark_seconds-warmup_seconds
   as it is currently.

There is a comment above IsCancelled that says when it is ok to use this.

We suffered a bug whereby doing a follow-up write to another write could
resurrect a deleted stream, causing all sorts of crash.

Fix: when a stream becomes writable (vs when we start writing) take a
ref on the stream, and only relinquish it once we're done writing.

could be an arbitrary amount of time between when the cancel is tried
and actually observable.

streaming API. In particular, Finish should not be called until the client
is sure that there is no more message to be read (as documented in
the comments for ClientStreamingInterface::Finish)