[FLINK-3787] Yarn client does not report unfulfillable container constraints

• Type: Improvement
• Status: Resolved
• Priority: Minor
• Resolution: Duplicate
• Affects Version/s: 1.1.0
• Fix Version/s: 1.1.4, 1.2.0
• Component/s: Deployment / YARN
• Assignee: Roman Maier
• Reporter: Till Rohrmann
• Created: 19/Apr/16 12:45
• Updated: 19/Dec/16 12:46
• Resolved: 28/Nov/16 10:19

Description

If the number of virtual cores for a Yarn container is not fulfillable, then the TaskManager won't be started. This is only reported in the logs but not in the FlinkYarnClient. Thus, the user will see a started JobManager with no connected TaskManagers. Since the log aggregation is only available after the Yarn job has been stopped, there is no easy way for the user to detect what's going on.

This problem is aggravated by the fact that the number of virtual cores is coupled to the number of slots if no explicit value has been set for the virtual cores. Therefore, it might happen that the Yarn deployment fails because of the virtual cores even though the user has never set a value for them (the user might even not know about the virtual cores).

I think it would be good to check if the virtual cores constraint is fulfillable. If not, then the user should receive a clear message that the Flink cluster cannot be deployed (similar to the memory constraints).

Issue Links

• duplicates: FLINK-4998 ResourceManager fails when num task slots > Yarn vcores