Details
Major Description
In https://issues.apache.org/jira/browse/KAFKA-12193, we changed the DNS resolution behavior for clients to re-resolve DNS after disconnecting from a broker, rather than wait until we iterated over all addresses from a given resolution. This is useful when the IP addresses have changed between the connection and disconnection.
However, with the behavior change, this does mean that clients could potentially reconnect immediately to the same IP they just disconnected from, if the IPs have not changed. In cases where the disconnection happened because that IP was unhealthy (such as a case where a load balancer has instances in multiple availability zones and one zone is unhealthy, or a case where an intermediate component in the network path is going through a rolling restart), this will delay the client successfully reconnecting. To address this, clients should remember the IP they just disconnected from and skip that IP when reconnecting, as long as the address resolved to multiple addresses.