Brooklin takes a long time to recover from errors in the destination kafka cluster, up to a complete halt of replication

Subject of the issue

Brooklin takes long time to recover from errors in the destination cluster. Sometimes multiple cycles of rebalances and complete halt of replication during that time.

I have run some tests, they are documented here: https://github.com/AppsFlyer/kafka-mirror-tester/blob/master/results-brooklin.md

Your environment

• Operating System 4.9.0-11-amd64 #1 SMP Debian 4.9.189-3+deb9u1 (2019-09-20) x86_64 Linux
• Brooklin version 1.0.2
• Java version openjdk version "1.8.0_212"
• Kafka version 2.1.0
• ZooKeeper version 3.4.10

Steps to reproduce

As described in this document https://github.com/AppsFlyer/kafka-mirror-tester/blob/master/results-brooklin.md, there are several scenarios in which that may happen.
One such scenario is to restart a broker at the destination cluster.
This results in errors for as long as the broker is down, which is understandable. But even long after the broker is back up - brooklin continues to err, up to a complete halt of replication to the entire cluster (not only to that failed broker).

Expected behaviour

We expect brooklin to recover gracefully and not halt replication during the rebalance cycle.
We expect to see just a single, hopefully short, rebalance, instead we multiple cycles that sometimes take quite long (10-15 minutes).

Actual behaviour

Brooklin takes a long time (10-15 minutes, sometimes more) to recover. During that time we see cycles of replication and then a complete halt of replication and then again, replication and then again a halt.

Start with the broker-restart scenario and the documented results in the linked kafka-mirror-tester report. Reproduce the destination-cluster failure and trace the rebalance and replication behavior while the broker is unavailable and after it returns. Done means recovery is graceful, with one short rebalance and no repeated replication halts.