[YARN-10787] Queue submit ACL check is wrong when CS queue is ambiguous

• Type: Bug
• Status: Patch Available
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 3.4.0
• Assignee: Gergely Pollák
• Reporter: Szilard Nemeth
• Created: 25/May/21 15:52
• Updated: 16/Jan/24 07:56

Description

Let's suppose we have a Capacity Scheduler configuration with 2 or more leaf queues with the same name in the queue hierarchy. That's what we call an ambiguous queue name.
Let's also enable ACL checks and define acl_submit_applications / acl_administer_queue configs with the correct value, adding the username to the ACL value there.

Here's a minimalistic YARN + CS config:

1. YARN config snippet:

<property><name>yarn.acl.enable</name><value>true</value>

2. CS config snippet:

<property>
	<name>yarn.scheduler.capacity.root.someparent1.queues</name>
	<value>anyotherqueue1,somequeue,anyotherqueue2</value>
</property>
<property>
	<name>yarn.scheduler.capacity.root.someparent2.queues</name>
	<value>anyotherqueue3,somequeue,anyotherqueue4</value>
</property>
<property>
	<name>yarn.scheduler.capacity.root.someparent1.somequeue.acl_submit_applications</name>
	<value>someuser1 </value>
</property>
<property>
	<name>yarn.scheduler.capacity.root.someparent2.somequeue.acl_submit_applications</name>
	<value>someuser1 </value>
</property>
<property>
	<name>yarn.scheduler.capacity.root.someparent1.somequeue.acl_administer_queue</name>
	<value>someuser1 </value>
</property>
<property>
	<name>yarn.scheduler.capacity.root.someparent2.somequeue.acl_administer_queue</name>
	<value>someuser1 </value>
</property>

So in this case, we have an ambiguous queue named "somequeue" under 2 different paths:

• root.someparent1.somequeue
• root.someparent2.somequeue

When a user submits an application correctly with the full queue path e.g. root.someparent1.somequeue, YARN will still fail to place the application to that queue and will use the short name in case ACL checking is enabled.

3. LOG SNIPPET

2021-05-20 22:04:32,031 DEBUG org.apache.hadoop.yarn.server.resourcemanager.placement.CSMappingPlacementRule: Placement final result 'root.someparent1.somequeue' for application 'application_1621540945412_0001'
 2021-05-20 22:04:32,031 INFO org.apache.hadoop.yarn.server.resourcemanager.RMAppManager: Placed application with ID application_1621540945412_0001 in queue: somequeue, original submission queue was: root.someparent1.somequeue
 2021-05-20 22:04:32,031 ERROR org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.CapacityScheduler: Ambiguous queue reference: somequeue please use full queue path instead.
 2021-05-20 22:04:32,031 INFO org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.CapacityScheduler: Application 'application_1621540945412_0001' is submitted without priority hence considering default queue/cluster priority: 0
 2021-05-20 22:04:32,032 INFO org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.CapacityScheduler: Priority '0' is acceptable in queue : somequeue for application: application_1621540945412_0001
 2021-05-20 22:04:32,993 INFO org.apache.hadoop.yarn.server.resourcemanager.ClientRMService: Exception in submitting application_1621540945412_0001
 org.apache.hadoop.yarn.exceptions.YarnException: org.apache.hadoop.security.AccessControlException: User someuser1 does not have permission to submit application_1621540945412_0001 to queue somequeue
 at org.apache.hadoop.yarn.ipc.RPCUtil.getRemoteException(RPCUtil.java:38)

4. FULL STACKTRACE:

 org.apache.hadoop.yarn.exceptions.YarnException: org.apache.hadoop.security.AccessControlException: User someuser1 does not have permission to submit application_1621540945412_0001 to queue somequeue
	at org.apache.hadoop.yarn.ipc.RPCUtil.getRemoteException(RPCUtil.java:38)
	at org.apache.hadoop.yarn.server.resourcemanager.RMAppManager.createAndPopulateNewRMApp(RMAppManager.java:433)
	at org.apache.hadoop.yarn.server.resourcemanager.RMAppManager.submitApplication(RMAppManager.java:330)
	at org.apache.hadoop.yarn.server.resourcemanager.ClientRMService.submitApplication(ClientRMService.java:650)
	at org.apache.hadoop.yarn.api.impl.pb.service.ApplicationClientProtocolPBServiceImpl.submitApplication(ApplicationClientProtocolPBServiceImpl.java:277)
	at org.apache.hadoop.yarn.proto.ApplicationClientProtocol$ApplicationClientProtocolService$2.callBlockingMethod(ApplicationClientProtocol.java:563)
	at org.apache.hadoop.ipc.ProtobufRpcEngine$Server$ProtoBufRpcInvoker.call(ProtobufRpcEngine.java:528)
	at org.apache.hadoop.ipc.RPC$Server.call(RPC.java:1070)
	at org.apache.hadoop.ipc.Server$RpcCall.run(Server.java:989)
	at org.apache.hadoop.ipc.Server$RpcCall.run(Server.java:917)
	at java.security.AccessController.doPrivileged(Native Method)
	at javax.security.auth.Subject.doAs(Subject.java:422)
	at org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:1898)
	at org.apache.hadoop.ipc.Server$Handler.run(Server.java:2894)
Caused by: org.apache.hadoop.security.AccessControlException: User someuser1 does not have permission to submit application_1621540945412_0001 to queue somequeue
	at org.apache.hadoop.yarn.server.resourcemanager.RMAppManager.createAndPopulateNewRMApp(RMAppManager.java:436)
	... 12 more
2021-05-20 22:04:32,994 WARN org.apache.hadoop.yarn.server.resourcemanager.RMAuditLogger: USER=someuser1	IP=172.17.61.133	OPERATION=Submit Application Request	TARGET=ClientRMService	RESULT=FAILURE	DESCRIPTION=Exception in submitting application	PERMISSIONS=org.apache.hadoop.security.AccessControlException: User someuser1 does not have permission to submit application_1621540945412_0001 to queue somequeue	APPID=application_1621540945412_0001	QUEUENAME=somequeue

DETAILS:

1. The whole thing happens in RMAppManager#createAndPopulateNewRMApp:
Class / method: org.apache.hadoop.yarn.server.resourcemanager.RMAppManager#createAndPopulateNewRMApp
LINK

2. RMAppManager#copyPlacementQueueToSubmissionContext is called for applications that are new, meaning we are not recovering, an application is submitted in a normal way:
Class / method: org.apache.hadoop.yarn.server.resourcemanager.RMAppManager#copyPlacementQueueToSubmissionContext

Called at

Method link

The problem is that copyPlacementQueueToSubmissionContext sets the queue of context (ApplicationSubmissionContext object) from placementContext.getQueue (ApplicationPlacementContext object). If placementcontext holds the queue name in the short form, this will override the default submission queue value, let's suppose it was the full queue path.
An example of a generated log from this method:

 2021-05-20 22:04:32,031 INFO org.apache.hadoop.yarn.server.resourcemanager.RMAppManager: Placed application with ID application_1621540945412_0001 in queue: somequeue, original submission queue was: root.someparent1.somequeue

3. The problematic code block is here: Code block

3.1 First, the short queuename will be gathered from submissionContext, as it was overridden by 'copyPlacementQueueToSubmissionContext': Link
This is a bad design, as here we are relying on the fact that the queue name was overridden in the submission context object.

3.2 Since the queue name will be in the short form and it's ambiguous, the call to scheduler.getQueue() will return null, as it's implemented like this by design: If the queue name is ambiguous, it returns null.

3.3 The condition of checking if csqueue is null AND placementContext is not null will evaluate to true here

3.4. The Parent queue will be queried from CS by the parent queue name of the placement context: Link

3.5 Finally, the ACL check fails as csqueue is the queue object of the parent queue of the queue 'root.someparent1.somequeue' which will be the queue: 'root.someparent1'.
In this case, the user don't have a submission ACL set for the parent queue, but the leaf queue so the ACL check fails.

LIST OF THINGS TO FIX / DO:

• Add a unit testcase that replicates the above config and the issue.
• Rename copyPlacementQueueToSubmissionContext: This method not really copies anything, it simply overrides the queue value.

• Add Debug log to print csqueue object before the authorization code: Auth code block

• Fix log messages: As 'copyPlacementQueueToSubmissionContext' overrides (not copies) the original queue name with the queue name from the PlacementContext, all calls to submissionContext.getQueue() will return the short queue name. This results in very misleading log messages as well, including the exception message itself:

   org.apache.hadoop.yarn.exceptions.YarnException: org.apache.hadoop.security.AccessControlException: User someuser1 does not have permission to submit application_1621540945412_0001 to queue somequeue
  

  All log messages should print the original submission queue, if possible.

• Actual code fix for the issue: Use full queue path to get the queue object.
  Again, this is the code block where the fix should happen: LINK

'queueName' should have the value set from: org.apache.hadoop.yarn.server.resourcemanager.placement.ApplicationPlacementContext#getFullQueuePath.

The equivalent of this in the linked code block:

placementContext.getFullQueuePath()

This should happen only if placementContext is not null.

LONG TERM FIX:

Investigate if it's possible to eliminate copyPlacementQueueToSubmissionContext.
This could introduce nasty backward incompatible issues with recovery, so it should be thought through really carefully.

Issue Links

• relates to: YARN-10798 Enhancements in RMAppManager: createAndPopulateNewRMApp and copyPlacementQueueToSubmissionContext
• relates to: YARN-10799 Eliminate queue name replacement in ApplicationSubmissionContext based on placement context