[KAFKA-7765] IdleExpiryManager should not passively close socket used by controller

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 2.1.0
• Component/s: network
• Assignee: Unassigned
• Reporter: huxihx
• Created: 21/Dec/18 08:56
• Updated: 27/Dec/18 09:46

Description

Currently, controller creates sockets for every living brokers without idle timeout. However, processor threads from other brokers still could close these sockets if no requests flow through them within `connections.max.idle.ms`.

Lots of CLOSE_WAITs were left when those sockets were closed by remote peer since controller's RequestSendThread will not check if they are closed by peer.

I think we need to figure out a way to record which channels should be maintained and have them excluded by IdleExpiryManager. A naive method is to augment KafkaChannel, making it have a field indicating whether this channel should be kept alive.

Does it make any sense?