[YARN-7720] Race condition between second app attempt and UAM timeout when first attempt node is down

• Type: Sub-task
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 3.4.0
• Fix Version/s: 3.4.0
• Component/s: federation
• Labels: pull-request-available
• Target Version/s: 3.4.0
• Hadoop Flags: Reviewed
• Assignee: Shilun Fan
• Reporter: Botong Huang
• Created: 09/Jan/18 01:19
• Updated: 12/Feb/24 06:49
• Resolved: 29/May/23 17:37

Description

In Federation, multiple attempts of an application share the same UAM in each secondary sub-cluster. When first attempt fails, we reply on the fact that secondary RM won't kill the existing UAM before the AM heartbeat timeout (default at 10 min). When second attempt comes up in the home sub-cluster, it will pick up the UAM token from Yarn Registry and resume the UAM heartbeat to secondary RMs.

The default heartbeat timeout for NM and AM are both 10 mins. The problem is that when the first attempt node goes down or out of connection, only after 10 mins will the home RM mark the first attempt as failed, and then schedule the 2nd attempt in some other node. By then the UAMs in secondaries are already timing out, and they might not survive until the second attempt comes up.

Issue Links

• links to: GitHub Pull Request #5672