Update: The solution is to disable vMotion on the VM where Kafka is deployed.

It seems that whenever vMotion perform a live migration, Kafka will be impacted and will result in broken pipe error. According to this link, Confluent also recommends to disable the vMotion as it can cause cluster outage.

After disabling the vMotion, my Kafka cluster never encounters any broken pipe issues again. This still doesn't answer why it won't automatically rejoin the cluster after encounter broken pipe error, but at least it does solve the broken pipe error.