[FLINK-10866] Queryable state can prevent cluster from starting

• Type: Improvement
• Status: Closed
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 1.5.5, 1.6.2, 1.7.0
• Fix Version/s: 1.8.0
• Component/s: Runtime / Task
• Labels: pull-request-available
• Assignee: BoWang
• Reporter: Till Rohrmann
• Created: 13/Nov/18 14:35
• Updated: 10/Jan/19 10:13
• Resolved: 10/Jan/19 10:13

Description

The KvStateServerImpl can currently prevent the TaskExecutor from starting.

Currently, the QS server starts per default on port 9067. If this port is not free, then it fails and stops the whole initialization of the TaskExecutor. I think the QS server should not stop the TaskExecutor from starting.

We should at least change the default port to 0 to avoid port conflicts. However, this will break all setups which don't explicitly set the QS port because now it either needs to be setup or extracted from the logs.

Additionally, we should think about whether a QS server startup failure should lead to a TaskExecutor failure or simply be logged. Both approaches have pros and cons. Currently, a failing QS server will also affect users which don't want to use QS. If we tolerate failures in the QS server, then a user who wants to use QS might run into problems with state not being reachable.

Issue Links

• links to: GitHub Pull Request #7412