[KAFKA-2677] Coordinator disconnects not propagated to new consumer

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 0.9.0.0
• Assignee: Jason Gustafson
• Reporter: Jason Gustafson
• Created: 20/Oct/15 21:26
• Updated: 26/Jan/17 03:20
• Resolved: 27/Oct/15 23:19

Description

Currently, disconnects by the coordinator are not always seen by the consumer. This can result in a long delay after the old coordinator has shutdown or failed before the consumer knows that it needs to find the new coordinator. The NetworkClient makes socket disconnects available to users in two ways:

1. through a flag in the ClientResponse object for requests pending when the disconnect occurred, and
2. through the connectionFailed() method.

The first method clearly cannot be depended on since it only helps when a request is pending, which is relatively rare for the connection with the coordinator. Instead, we can probably use the second method with a little rework of ConsumerNetworkClient to check for failed connections immediately after returning from poll().

Issue Links

• is related to: KAFKA-2615 Poll() method is broken wrt time