[YARN-7377] Duplicate Containers allocated for Long-Running Application after NM lost and restart and RM restart

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 3.0.0-alpha3
• Component/s: applications, nodemanager, RM, yarn
• Labels: patch

Environment

RM recovery and NM recovery enabled；
Spark streaming application, a long-running application on yarn

• Assignee: Unassigned
• Reporter: rangjiaheng
• Created: 21/Oct/17 04:40
• Updated: 21/Oct/17 05:48

Description

Case:
A Spark streaming application named app1 running on yarn for a long time; app1 has 3 containers in total, one of them named c1 runs on a NM named nm1;

1. The NM named nm1 was lost for some reason, but the containers on it runs well;

2. 10 minutes later, RM lost this NM because of no heartbeats received; so RM tells app1's AM that a container of app1 was failed because of NM lost, so app1's AM killed that container through RPC and then request a new container named c2 from RM, which is duplicate to c1;

3. Administrator found nm1 lost, so he restart it; since NM's recovery was enabled, NM restore all the containers including container c1, but now c1's status is 'DONE';
A bug here: nm1 will list this container c1 in webui forever;

4. RM restart for some reason; since RM's recovery was enabled, RM restore all the apps including app1, and all the NM need re-register to RM; However, when nm1 registers to RM, RM found the container c1's status was DONE, so RM tells app1's AM that a container of app1 was complete, since spark streaming application has fixed number of containers, so AM request a new container named c3 from RM, which is duplicate to c1.

A bug here:
Now, app1 has 4 containers in total, while c2 and c3 were the same.