[KAFKA-4841] NetworkClient should only consider a connection to be fail after attempt to connect

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 0.11.0.0
• Assignee: Dong Lin
• Reporter: Dong Lin
• Created: 05/Mar/17 10:21
• Updated: 08/Mar/17 01:23
• Resolved: 08/Mar/17 01:23

Description

KAFKA-4820 allows new request to be enqueued to unsent by user thread while some other thread does poll(...). This causes problem in the following scenario:

• Thread A calls poll(...) and is blocked on select(...)
• Thread B enqueues a request into unsent of ConsumerNetworkClient for node N
• Thread A calls checkDisconnects(now) -> client.connectionFailed(N)

Because no attempts have been made to connection to node N yet, there is no state for node N and connectionFailed(N) would throw exception. Note that this problem only occurs when one thread is able to enqueue requests while another thread is in the process of `poll(...)`

The solution is to only consider a connection has failed if attempts have been made to connect to this node AND the connection state is DISCONNECTED.

Issue Links

• relates to: KAFKA-4842 Streams integration tests occasionally fail with connection error
• links to: GitHub Pull Request #2641