Are Java and Go doing the same? @ejona86 @dfawley

Reviewed 2 of 2 files at r1.
Reviewable status: complete! all files reviewed, all discussions resolved (waiting on @apolcyn)

src/core/ext/filters/client_channel/lb_policy/round_robin/round_robin.cc, line 388 at r1 (raw file):

do not change the aggregation state 

Nit: might be clearer to say something like "do not report state changes" so that the "not changing state" behavior sounds more like a subchannel limitation other than aggregation state limitation for every subchannel.

test/cpp/end2end/client_lb_end2end_test.cc, line 1294 at r1 (raw file):

  EXPECT_TRUE(WaitForChannelReady(channel.get()));
  // Now kill the servers.  The channel should transition to TRANSIENT_FAILURE.
  // TODO(roth): This test should ideally check that even when the

This is unfortunate but I can't think of any easy way to tune the connection time either.

The impetus for this PR was a suggestion from Eric and Doug. I don't think they've made this change in Java or Go yet, but we'll probably do this as part of the xDS work, since we're thinking that we'll want to use this same aggregation algorithm across priorities in the xds policy.

Reviewable status: complete! all files reviewed, all discussions resolved (waiting on @apolcyn)

src/core/ext/filters/client_channel/lb_policy/round_robin/round_robin.cc, line 388 at r1 (raw file):

I've updated the comment to hopefully make this clearer.

Reviewed 1 of 1 files at r2.
Reviewable status: complete! all files reviewed, all discussions resolved (waiting on @apolcyn)

Reviewed 1 of 1 files at r3.
Reviewable status: complete! all files reviewed, all discussions resolved (waiting on @apolcyn)