[FLINK-7469] Handle slot requests occuring before RM registration completes

• Type: Sub-task
• Status: Resolved
• Priority: Minor
• Resolution: Duplicate
• Fix Version/s: 1.5.0
• Component/s: Runtime / Coordination
• Assignee: Unassigned
• Reporter: Eron Wright
• Created: 17/Aug/17 17:45
• Updated: 28/Feb/19 14:12
• Resolved: 13/Jul/18 06:32

Description

Description
Occasionally the TM-to-RM registration ask times out, causing the TM to pause registration for 10 seconds. Meanwhile the registration may actually have succeeded in the RM. Slot requests may then arrive at the TM while RM registration is incomplete.

The current behavior appears to be that the TM honors the slot request. Please determine whether this is a feature or a bug. If a feature, maybe a slot request should implicitly complete the registration.

Example
See attached a log showing a certain TM exhibiting the described behavior. The RM launched 12 TMs in parallel, evidently causing the RM to sluggishly respond to a couple of the TM registration requests. From the logs we see that '00012' and '00003' experienced a registration timeout but accepted a slot request anyway.

Issue Links

• duplicates: FLINK-9427 Cannot download from BlobServer, because the server address is unknown.