[jira] [Commented] (YARN-371) Consolidate resource requests in AM-RM heartbeat
[ https://issues.apache.org/jira/browse/YARN-371?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13570109#comment-13570109 ] Tom White commented on YARN-371: Looks like there's a misunderstanding here - Sandy talks about _reducing_ the memory requirements of the RM. If I understand the proposal correctly, the number of resource request objects sent by the AM in MR would be reduced from five (three node-local, one rack-local, one ANY) to one resource request with an array of locations (host names) of length five. BTW Arun, immediately vetoing an issue in the first comment is not conducive to a balanced discussion! Consolidate resource requests in AM-RM heartbeat Key: YARN-371 URL: https://issues.apache.org/jira/browse/YARN-371 Project: Hadoop YARN Issue Type: Improvement Components: api, resourcemanager, scheduler Affects Versions: 2.0.2-alpha Reporter: Sandy Ryza Assignee: Sandy Ryza Each AMRM heartbeat consists of a list of resource requests. Currently, each resource request consists of a container count, a resource vector, and a location, which may be a node, a rack, or *. When an application wishes to request a task run in multiple localtions, it must issue a request for each location. This means that for a node-local task, it must issue three requests, one at the node-level, one at the rack-level, and one with * (any). These requests are not linked with each other, so when a container is allocated for one of them, the RM has no way of knowing which others to get rid of. When a node-local container is allocated, this is handled by decrementing the number of requests on that node's rack and in *. But when the scheduler allocates a task with a node-local request on its rack, the request on the node is left there. This can cause delay-scheduling to try to assign a container on a node that nobody cares about anymore. Additionally, unless I am missing something, the current model does not allow requests for containers only on a specific node or specific rack. While this is not a use case for MapReduce currently, it is conceivable that it might be something useful to support in the future, for example to schedule long-running services that persist state in a particular location, or for applications that generally care less about latency than data-locality. Lastly, the ability to understand which requests are for the same task will possibly allow future schedulers to make more intelligent scheduling decisions, as well as permit a more exact understanding of request load. I would propose the tweak of allowing a single ResourceRequest to encapsulate all the location information for a task. So instead of just a single location, a ResourceRequest would contain an array of locations, including nodes that it would be happy with, racks that it would be happy with, and possibly *. Side effects of this change would be a reduction in the amount of data that needs to be transferred in a heartbeat, as well in as the RM's memory footprint, becaused what used to be different requests for the same task are now able to share some common data. While this change breaks compatibility, if it is going to happen, it makes sense to do it now, before YARN becomes beta. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (YARN-374) Job History Server doesn't show jobs which killed by ClientRMProtocol.forceKillApplication
[ https://issues.apache.org/jira/browse/YARN-374?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13570285#comment-13570285 ] Thomas Graves commented on YARN-374: At a high level the job history server is currently a mapreduce specific component. So this behavior is currently expected. yarn/RM doesn't know about any history server, so if you force kill something through the RM it has no knowledge on how to handle the history, it simply force kills the application. There is another jira that is looking at making the history server generic that would help with this issue - see YARN-321 Job History Server doesn't show jobs which killed by ClientRMProtocol.forceKillApplication -- Key: YARN-374 URL: https://issues.apache.org/jira/browse/YARN-374 Project: Hadoop YARN Issue Type: Bug Components: client, resourcemanager Affects Versions: 2.0.1-alpha Reporter: nemon lou After i kill a app by typing bin/yarn rmadmin app -kill APP_ID, no job info is kept on JHS web page. However, when i kill a job by typing bin/mapred job -kill JOB_ID , i can see a killed job left on JHS. Some hive users are confused by that their jobs been killed but nothing left on JHS ,and killed app's info on RM web page is not enough.(They kill job by clientRMProtocol) -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (YARN-371) Consolidate resource requests in AM-RM heartbeat
[ https://issues.apache.org/jira/browse/YARN-371?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13570314#comment-13570314 ] Arun C Murthy commented on YARN-371: {quote} Looks like there's a misunderstanding here - Sandy talks about reducing the memory requirements of the RM. If I understand the proposal correctly, the number of resource request objects sent by the AM in MR would be reduced from five (three node-local, one rack-local, one ANY) to one resource request with an array of locations (host names) of length five. {quote} Please read my explanation again. This change is *explicitly* against the design goals of YARN ResourceManager and would increase memory requirements of RM by a couple of orders of magnitude. Hadoop MR applications, routinely, have 100K+ tasks. The proposed change in this jira would require 100K+ resource-requests (one per task). Currently, in YARN, that can be expressed in O(nodes + racks + 1) resource-requests, which is ~O(5000) on even the largest clusters known today. So, in effect, this change would be a significant regression and result in 100,000 resource-requests v/s ~5000 needed today. bq. BTW Arun, immediately vetoing an issue in the first comment is not conducive to a balanced discussion! Tom - You can read it as a veto, or you can read it as *I strongly disagree since this is against the goals of the project and a significant regression*. IAC, we should allow for people's communication style... and keep discussions technical - I'd appreciate that. Consolidate resource requests in AM-RM heartbeat Key: YARN-371 URL: https://issues.apache.org/jira/browse/YARN-371 Project: Hadoop YARN Issue Type: Improvement Components: api, resourcemanager, scheduler Affects Versions: 2.0.2-alpha Reporter: Sandy Ryza Assignee: Sandy Ryza Each AMRM heartbeat consists of a list of resource requests. Currently, each resource request consists of a container count, a resource vector, and a location, which may be a node, a rack, or *. When an application wishes to request a task run in multiple localtions, it must issue a request for each location. This means that for a node-local task, it must issue three requests, one at the node-level, one at the rack-level, and one with * (any). These requests are not linked with each other, so when a container is allocated for one of them, the RM has no way of knowing which others to get rid of. When a node-local container is allocated, this is handled by decrementing the number of requests on that node's rack and in *. But when the scheduler allocates a task with a node-local request on its rack, the request on the node is left there. This can cause delay-scheduling to try to assign a container on a node that nobody cares about anymore. Additionally, unless I am missing something, the current model does not allow requests for containers only on a specific node or specific rack. While this is not a use case for MapReduce currently, it is conceivable that it might be something useful to support in the future, for example to schedule long-running services that persist state in a particular location, or for applications that generally care less about latency than data-locality. Lastly, the ability to understand which requests are for the same task will possibly allow future schedulers to make more intelligent scheduling decisions, as well as permit a more exact understanding of request load. I would propose the tweak of allowing a single ResourceRequest to encapsulate all the location information for a task. So instead of just a single location, a ResourceRequest would contain an array of locations, including nodes that it would be happy with, racks that it would be happy with, and possibly *. Side effects of this change would be a reduction in the amount of data that needs to be transferred in a heartbeat, as well in as the RM's memory footprint, becaused what used to be different requests for the same task are now able to share some common data. While this change breaks compatibility, if it is going to happen, it makes sense to do it now, before YARN becomes beta. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (YARN-371) Consolidate resource requests in AM-RM heartbeat
[ https://issues.apache.org/jira/browse/YARN-371?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13570325#comment-13570325 ] Arun C Murthy commented on YARN-371: Sandy - please don't let this side discussion distract you, it's an individual style thing. I use -1 on grocery list discussions with my wife... unfortunately I don't have the luxury of vetos in that context! *smile* Anyway, there are other good discussion points such as 'allowing requests on a specific node/rack' which I have pondered about for a long while too. Maybe we can close this jira and open one for specific enhancements? In future, it would help if jira descriptions are short and propose a specific enhancement - this way we can debate solutions separately (maybe even on *-dev list). On the plus side, this way I can -1 a specific implementation proposal rather than the jira too... ;-) Consolidate resource requests in AM-RM heartbeat Key: YARN-371 URL: https://issues.apache.org/jira/browse/YARN-371 Project: Hadoop YARN Issue Type: Improvement Components: api, resourcemanager, scheduler Affects Versions: 2.0.2-alpha Reporter: Sandy Ryza Assignee: Sandy Ryza Each AMRM heartbeat consists of a list of resource requests. Currently, each resource request consists of a container count, a resource vector, and a location, which may be a node, a rack, or *. When an application wishes to request a task run in multiple localtions, it must issue a request for each location. This means that for a node-local task, it must issue three requests, one at the node-level, one at the rack-level, and one with * (any). These requests are not linked with each other, so when a container is allocated for one of them, the RM has no way of knowing which others to get rid of. When a node-local container is allocated, this is handled by decrementing the number of requests on that node's rack and in *. But when the scheduler allocates a task with a node-local request on its rack, the request on the node is left there. This can cause delay-scheduling to try to assign a container on a node that nobody cares about anymore. Additionally, unless I am missing something, the current model does not allow requests for containers only on a specific node or specific rack. While this is not a use case for MapReduce currently, it is conceivable that it might be something useful to support in the future, for example to schedule long-running services that persist state in a particular location, or for applications that generally care less about latency than data-locality. Lastly, the ability to understand which requests are for the same task will possibly allow future schedulers to make more intelligent scheduling decisions, as well as permit a more exact understanding of request load. I would propose the tweak of allowing a single ResourceRequest to encapsulate all the location information for a task. So instead of just a single location, a ResourceRequest would contain an array of locations, including nodes that it would be happy with, racks that it would be happy with, and possibly *. Side effects of this change would be a reduction in the amount of data that needs to be transferred in a heartbeat, as well in as the RM's memory footprint, becaused what used to be different requests for the same task are now able to share some common data. While this change breaks compatibility, if it is going to happen, it makes sense to do it now, before YARN becomes beta. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (YARN-371) Consolidate resource requests in AM-RM heartbeat
[ https://issues.apache.org/jira/browse/YARN-371?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13570329#comment-13570329 ] Robert Joseph Evans commented on YARN-371: -- Tom just like Arun said the memory usage changes based off of the size of the cluster vs. the size of the request. The current approach is on the order of the size of the cluster where as the proposed approach is on the order of the number of desired containers. If I have a 100 node cluster and I am requesting 10 map tasks the size will be O(100 nodes + X racks + 1) possibly * 2 if reducers are included in it. What is more it is probably exactly the same size of request for 1 or even 1000 tasks. Where as the proposed approach would grow without bound as the number of tasks also increased. However, I also agree with Sandy that the current state compression is lossy and as such restricts what is possible in the scheduler. I would like to understand better what the size differences would be for various requests, both in memory and also over the wire. It seems conceivable to me that if the size difference is not too big, especially over the wire, we could allow the scheduler itself to decide on its in memory representation. This would allow for the Capacity Scheduler to keep its current layout and allow for others to experiment with more advanced scheduling options. Different groups could decide which scheduler best fits their needs and workload. If the size is significantly larger I would like to see hard numbers about how much better/worse it makes specific use cases. I am also very concerned about adding too much complexity to the scheduler. We have run into issues where the RM will get very far behind in scheduling because it is trying to do a lot already and eventually OOM as its event queue grows too large. I also don't want to change the scheduler protocol too much without first understanding how that new protocol would impact other potential scheduling features. There are a number of other computing patterns that could benefit from specific scheduler support. Things like gang scheduling where you need all of the containers at once or none of them can make any progress, or where you want all of the containers to be physically close to one another because they are very I/O intensive, but you don't really care where exactly they are. Or even something like HBase where you essentially want one process on every single node with no duplicates. Do the proposed changes make these uses case trivially simple, or do they require a lot of support on the AM to implement them? Consolidate resource requests in AM-RM heartbeat Key: YARN-371 URL: https://issues.apache.org/jira/browse/YARN-371 Project: Hadoop YARN Issue Type: Improvement Components: api, resourcemanager, scheduler Affects Versions: 2.0.2-alpha Reporter: Sandy Ryza Assignee: Sandy Ryza Each AMRM heartbeat consists of a list of resource requests. Currently, each resource request consists of a container count, a resource vector, and a location, which may be a node, a rack, or *. When an application wishes to request a task run in multiple localtions, it must issue a request for each location. This means that for a node-local task, it must issue three requests, one at the node-level, one at the rack-level, and one with * (any). These requests are not linked with each other, so when a container is allocated for one of them, the RM has no way of knowing which others to get rid of. When a node-local container is allocated, this is handled by decrementing the number of requests on that node's rack and in *. But when the scheduler allocates a task with a node-local request on its rack, the request on the node is left there. This can cause delay-scheduling to try to assign a container on a node that nobody cares about anymore. Additionally, unless I am missing something, the current model does not allow requests for containers only on a specific node or specific rack. While this is not a use case for MapReduce currently, it is conceivable that it might be something useful to support in the future, for example to schedule long-running services that persist state in a particular location, or for applications that generally care less about latency than data-locality. Lastly, the ability to understand which requests are for the same task will possibly allow future schedulers to make more intelligent scheduling decisions, as well as permit a more exact understanding of request load. I would propose the tweak of allowing a single ResourceRequest to encapsulate all the location information for a task. So instead of just a single location, a
[jira] [Created] (YARN-375) FIFO scheduler may crash due to bugg app
Eli Collins created YARN-375: Summary: FIFO scheduler may crash due to bugg app Key: YARN-375 URL: https://issues.apache.org/jira/browse/YARN-375 Project: Hadoop YARN Issue Type: Bug Affects Versions: 2.0.0-alpha Reporter: Eli Collins Priority: Critical The following code should check for a 0 return value rather than crash! {code} int availableContainers = node.getAvailableResource().getMemory() / capability.getMemory(); // TODO: A buggy // application // with this // zero would // crash the // scheduler. {code} -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (YARN-370) CapacityScheduler app submission fails when min alloc size not multiple of AM size
[ https://issues.apache.org/jira/browse/YARN-370?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13570514#comment-13570514 ] Thomas Graves commented on YARN-370: Sorry the only other thing I can think of that would matter is having security on. I had security on and the code that throws the exception is looking at the Token, so if you don't have security on you probably won't see it. Other then that it was running any simple job - sleep, wordcount. CapacityScheduler app submission fails when min alloc size not multiple of AM size -- Key: YARN-370 URL: https://issues.apache.org/jira/browse/YARN-370 Project: Hadoop YARN Issue Type: Bug Components: capacityscheduler Affects Versions: 2.0.3-alpha Reporter: Thomas Graves Assignee: Zhijie Shen Priority: Blocker I was running 2.0.3-SNAPSHOT with the capacity scheduler configured with minimum allocation size 1G. The AM size was set to 1.5G. I didn't specify resource calculator so it was using DefaultResourceCalculator. The am launch failed with the error below: Application application_1359688216672_0001 failed 1 times due to Error launching appattempt_1359688216672_0001_01. Got exception: RemoteTrace: at LocalTrace: org.apache.hadoop.yarn.exceptions.impl.pb.YarnRemoteExceptionPBImpl: RemoteTrace: at LocalTrace: org.apache.hadoop.yarn.exceptions.impl.pb.YarnRemoteExceptionPBImpl: Unauthorized request to start container. Expected resource memory:2048, vCores:1 but found memory:1536, vCores:1 at org.apache.hadoop.yarn.factories.impl.pb.YarnRemoteExceptionFactoryPBImpl.createYarnRemoteException(YarnRemoteExceptionFactoryPBImpl.java:39) at org.apache.hadoop.yarn.ipc.RPCUtil.getRemoteException(RPCUtil.java:47) at org.apache.hadoop.yarn.server.nodemanager.containermanager.ContainerManagerImpl.authorizeRequest(ContainerManagerImpl.java:383) at org.apache.hadoop.yarn.server.nodemanager.containermanager.ContainerManagerImpl.startContainer(ContainerManagerImpl.java:400) at org.apache.hadoop.yarn.api.impl.pb.service.ContainerManagerPBServiceImpl.startContainer(ContainerManagerPBServiceImpl.java:68) at org.apache.hadoop.yarn.proto.ContainerManager$ContainerManagerService$2.callBlockingMethod(ContainerManager.java:83) at org.apache.hadoop.ipc.ProtobufRpcEngine$Server$ProtoBufRpcInvoker.call(ProtobufRpcEngine.java:454) at org.apache.hadoop.ipc.RPC$Server.call(RPC.java:1014) at org.apache.hadoop.ipc.Server$Handler$1.run(Server.java:1735) at org.apache.hadoop.ipc.Server$Handler$1.run(Server.java:1731) at java.security.AccessController.doPrivileged(Native Method) at javax.security.auth.Subject.doAs(Subject.java:415) at org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:1441) at org.apache.hadoop.ipc.Server$Handler.run(Server.java:1729) at sun.reflect.NativeConstructorAccessorImpl.newInstance0(Native Method) at sun.reflect.NativeConstructorAccessorImpl.newInstance(NativeConstructorAccessorImpl.java:57) at sun.reflect.DelegatingConstructorAccessorImpl.newInstance(DelegatingConstructorAccessorImpl.java:45) at java.lang.reflect.Constructor.newInstance(Constructor.java:525) at org.apache.hadoop.ipc.RemoteException.instantiateException(RemoteException.java:90) at org.apache.hadoop.ipc.RemoteException.unwrapRemoteException(RemoteException.java:57) at org.apache.hadoop.yarn.exceptions.impl.pb.YarnRemoteExceptionPBImpl.unwrapAndThrowException(YarnRemoteExceptionPBImpl.java:123) at org.apache.hadoop.yarn.api.impl.pb.client.ContainerManagerPBClientImpl.startContainer(ContainerManagerPBClientImpl.java:109) at org.apache.hadoop.yarn.server.resourcemanager.amlauncher.AMLauncher.launch(AMLauncher.java:111) at org.apache.hadoop.yarn.server.resourcemanager.amlauncher.AMLauncher.run(AMLauncher.java:255) at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1110) at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:603) at java.lang.Thread.run(Thread.java:722) . Failing the application. It looks like the launchcontext for the app didn't have the resources rounded up. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (YARN-375) FIFO scheduler may crash due to bugg app
[ https://issues.apache.org/jira/browse/YARN-375?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13570545#comment-13570545 ] Arun C Murthy commented on YARN-375: I believe this doesn't happen at all since there is a check upfront, but I'll double-check to make sure. FIFO scheduler may crash due to bugg app -- Key: YARN-375 URL: https://issues.apache.org/jira/browse/YARN-375 Project: Hadoop YARN Issue Type: Bug Affects Versions: 2.0.0-alpha Reporter: Eli Collins Assignee: Arun C Murthy Priority: Critical The following code should check for a 0 return value rather than crash! {code} int availableContainers = node.getAvailableResource().getMemory() / capability.getMemory(); // TODO: A buggy // application // with this // zero would // crash the // scheduler. {code} -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Created] (YARN-376) Apps that have completed can appear as RUNNING on the NM UI
Jason Lowe created YARN-376: --- Summary: Apps that have completed can appear as RUNNING on the NM UI Key: YARN-376 URL: https://issues.apache.org/jira/browse/YARN-376 Project: Hadoop YARN Issue Type: Bug Components: resourcemanager Affects Versions: 2.0.3-alpha, 0.23.6 Reporter: Jason Lowe On a busy cluster we've noticed a growing number of applications appear as RUNNING on a nodemanager web pages but the applications have long since finished. Looking at the NM logs, it appears the RM never told the nodemanager that the application had finished. This is also reflected in a jstack of the NM process, since many more log aggregation threads are running then one would expect from the number of actively running applications. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (YARN-376) Apps that have completed can appear as RUNNING on the NM UI
[ https://issues.apache.org/jira/browse/YARN-376?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13570568#comment-13570568 ] Jason Lowe commented on YARN-376: - There appears to be a race condition in the RM's handling of finished applications that may explain this. ResourceTrackerService is sending the list of finished applications to the node when the node heartbeats and then subsequently sending a status update event to the RMNodeImpl that corresponds to the node. The RMNodeImpl clears the entire list of finished applications once it has processed the status update. If an application completes *after* the ResourceTrackerService has asynchronously retrieved the list of finished applications but *before* the status update event is posted to the RMNodeImpl then the application will be added to then cleared from the list of finished applications before the ResourceTrackerService had a chance to notify the node of the completing application. Apps that have completed can appear as RUNNING on the NM UI --- Key: YARN-376 URL: https://issues.apache.org/jira/browse/YARN-376 Project: Hadoop YARN Issue Type: Bug Components: resourcemanager Affects Versions: 2.0.3-alpha, 0.23.6 Reporter: Jason Lowe On a busy cluster we've noticed a growing number of applications appear as RUNNING on a nodemanager web pages but the applications have long since finished. Looking at the NM logs, it appears the RM never told the nodemanager that the application had finished. This is also reflected in a jstack of the NM process, since many more log aggregation threads are running then one would expect from the number of actively running applications. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Updated] (YARN-5) Add support for FifoScheduler to schedule CPU along with memory.
[ https://issues.apache.org/jira/browse/YARN-5?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ] Eli Collins updated YARN-5: --- Issue Type: New Feature (was: Sub-task) Parent: (was: YARN-2) Add support for FifoScheduler to schedule CPU along with memory. Key: YARN-5 URL: https://issues.apache.org/jira/browse/YARN-5 Project: Hadoop YARN Issue Type: New Feature Reporter: Arun C Murthy Assignee: Arun C Murthy -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (YARN-371) Resource-centric compression in AM-RM protocol limits scheduling
[ https://issues.apache.org/jira/browse/YARN-371?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13570608#comment-13570608 ] Sandy Ryza commented on YARN-371: - Bobby, I believe that a task-centric request format is necessary for the cases you mention, but not entirely sufficient for all of them. All would likely require significant modifications to the scheduler. For the HBase case, each task request could simply be at a single node (no racks or *). For the case of applications that want containers located near each other, but don't care where, tasks could include a special location value that means try to put me near other tasks that share this value. I believe gang-scheduling would require a task-centric protocol as well, but would either require a flag that says the entire heartbeat should be gang-scheduled or a grouping of requests within a heartbeat into gangs. Resource-centric compression in AM-RM protocol limits scheduling Key: YARN-371 URL: https://issues.apache.org/jira/browse/YARN-371 Project: Hadoop YARN Issue Type: Improvement Components: api, resourcemanager, scheduler Affects Versions: 2.0.2-alpha Reporter: Sandy Ryza Assignee: Sandy Ryza Each AMRM heartbeat consists of a list of resource requests. Currently, each resource request consists of a container count, a resource vector, and a location, which may be a node, a rack, or *. When an application wishes to request a task run in multiple localtions, it must issue a request for each location. This means that for a node-local task, it must issue three requests, one at the node-level, one at the rack-level, and one with * (any). These requests are not linked with each other, so when a container is allocated for one of them, the RM has no way of knowing which others to get rid of. When a node-local container is allocated, this is handled by decrementing the number of requests on that node's rack and in *. But when the scheduler allocates a task with a node-local request on its rack, the request on the node is left there. This can cause delay-scheduling to try to assign a container on a node that nobody cares about anymore. Additionally, unless I am missing something, the current model does not allow requests for containers only on a specific node or specific rack. While this is not a use case for MapReduce currently, it is conceivable that it might be something useful to support in the future, for example to schedule long-running services that persist state in a particular location, or for applications that generally care less about latency than data-locality. Lastly, the ability to understand which requests are for the same task will possibly allow future schedulers to make more intelligent scheduling decisions, as well as permit a more exact understanding of request load. I would propose the tweak of allowing a single ResourceRequest to encapsulate all the location information for a task. So instead of just a single location, a ResourceRequest would contain an array of locations, including nodes that it would be happy with, racks that it would be happy with, and possibly *. Side effects of this change would be a reduction in the amount of data that needs to be transferred in a heartbeat, as well in as the RM's memory footprint, becaused what used to be different requests for the same task are now able to share some common data. While this change breaks compatibility, if it is going to happen, it makes sense to do it now, before YARN becomes beta. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (YARN-371) Resource-centric compression in AM-RM protocol limits scheduling
[ https://issues.apache.org/jira/browse/YARN-371?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13570619#comment-13570619 ] Robert Joseph Evans commented on YARN-371: -- I didn't really expect them to be trivial :). So I think that there may be some value in having a different protocol, but we need some hard numbers to be able to really make an informed decision. I would like to see the size of a request in the following table (both in memory size on the RM and size sent over the wire) ||nodes(down)/tasks(across)||1,000||10,000||100,000||500,000|| ||100|?|?|?|?| ||1,000|?|?|?|?| ||4,000|?|?|?|?| ||10,000|?|?|?|?| It would also be great to see in practice how bad is the scheduling problem where the wrong node is sent. Resource-centric compression in AM-RM protocol limits scheduling Key: YARN-371 URL: https://issues.apache.org/jira/browse/YARN-371 Project: Hadoop YARN Issue Type: Improvement Components: api, resourcemanager, scheduler Affects Versions: 2.0.2-alpha Reporter: Sandy Ryza Assignee: Sandy Ryza Each AMRM heartbeat consists of a list of resource requests. Currently, each resource request consists of a container count, a resource vector, and a location, which may be a node, a rack, or *. When an application wishes to request a task run in multiple localtions, it must issue a request for each location. This means that for a node-local task, it must issue three requests, one at the node-level, one at the rack-level, and one with * (any). These requests are not linked with each other, so when a container is allocated for one of them, the RM has no way of knowing which others to get rid of. When a node-local container is allocated, this is handled by decrementing the number of requests on that node's rack and in *. But when the scheduler allocates a task with a node-local request on its rack, the request on the node is left there. This can cause delay-scheduling to try to assign a container on a node that nobody cares about anymore. Additionally, unless I am missing something, the current model does not allow requests for containers only on a specific node or specific rack. While this is not a use case for MapReduce currently, it is conceivable that it might be something useful to support in the future, for example to schedule long-running services that persist state in a particular location, or for applications that generally care less about latency than data-locality. Lastly, the ability to understand which requests are for the same task will possibly allow future schedulers to make more intelligent scheduling decisions, as well as permit a more exact understanding of request load. I would propose the tweak of allowing a single ResourceRequest to encapsulate all the location information for a task. So instead of just a single location, a ResourceRequest would contain an array of locations, including nodes that it would be happy with, racks that it would be happy with, and possibly *. Side effects of this change would be a reduction in the amount of data that needs to be transferred in a heartbeat, as well in as the RM's memory footprint, becaused what used to be different requests for the same task are now able to share some common data. While this change breaks compatibility, if it is going to happen, it makes sense to do it now, before YARN becomes beta. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (YARN-371) Resource-centric compression in AM-RM protocol limits scheduling
[ https://issues.apache.org/jira/browse/YARN-371?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13570633#comment-13570633 ] Arun C Murthy commented on YARN-371: Jira isn't a particularly good medium to have a discussion of this kind - we should move this to yarn-dev@. I'm very wary of supporting multiple protocols (task-centric v/s resource-centric) to the point of being paranoid about it. Supporting multiple protocols or APIs is very expensive - look at the hardships we have had with mapred v/s mapreduce apis. The *task-centric* protocol in the current JobTracker is something we *know* doesn't work well at scale (cluster sizes, number of concurrent applications etc.), we need to remember that - I have lots of scars I don't want to re-open. Instead, we should focus on specific use-cases and debate how we can fix them in the context of a protocol which we know scales well as it stands. Resource-centric compression in AM-RM protocol limits scheduling Key: YARN-371 URL: https://issues.apache.org/jira/browse/YARN-371 Project: Hadoop YARN Issue Type: Improvement Components: api, resourcemanager, scheduler Affects Versions: 2.0.2-alpha Reporter: Sandy Ryza Assignee: Sandy Ryza Each AMRM heartbeat consists of a list of resource requests. Currently, each resource request consists of a container count, a resource vector, and a location, which may be a node, a rack, or *. When an application wishes to request a task run in multiple localtions, it must issue a request for each location. This means that for a node-local task, it must issue three requests, one at the node-level, one at the rack-level, and one with * (any). These requests are not linked with each other, so when a container is allocated for one of them, the RM has no way of knowing which others to get rid of. When a node-local container is allocated, this is handled by decrementing the number of requests on that node's rack and in *. But when the scheduler allocates a task with a node-local request on its rack, the request on the node is left there. This can cause delay-scheduling to try to assign a container on a node that nobody cares about anymore. Additionally, unless I am missing something, the current model does not allow requests for containers only on a specific node or specific rack. While this is not a use case for MapReduce currently, it is conceivable that it might be something useful to support in the future, for example to schedule long-running services that persist state in a particular location, or for applications that generally care less about latency than data-locality. Lastly, the ability to understand which requests are for the same task will possibly allow future schedulers to make more intelligent scheduling decisions, as well as permit a more exact understanding of request load. I would propose the tweak of allowing a single ResourceRequest to encapsulate all the location information for a task. So instead of just a single location, a ResourceRequest would contain an array of locations, including nodes that it would be happy with, racks that it would be happy with, and possibly *. Side effects of this change would be a reduction in the amount of data that needs to be transferred in a heartbeat, as well in as the RM's memory footprint, becaused what used to be different requests for the same task are now able to share some common data. While this change breaks compatibility, if it is going to happen, it makes sense to do it now, before YARN becomes beta. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Created] (YARN-377) Fix test failure for HADOOP-9252
Tsz Wo (Nicholas), SZE created YARN-377: --- Summary: Fix test failure for HADOOP-9252 Key: YARN-377 URL: https://issues.apache.org/jira/browse/YARN-377 Project: Hadoop YARN Issue Type: Bug Reporter: Tsz Wo (Nicholas), SZE Priority: Minor HADOOP-9252 slightly changes the format of some StringUtils outputs. It may cause test failures. Also, some methods was deprecated by HADOOP-9252. The use of them should be replaced with the new methods. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Updated] (YARN-377) Fix test failure for HADOOP-9252
[ https://issues.apache.org/jira/browse/YARN-377?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ] Tsz Wo (Nicholas), SZE updated YARN-377: Description: HADOOP-9252 slightly changes the format of some StringUtils outputs. It may cause test failures. Also, some methods were deprecated by HADOOP-9252. The use of them should be replaced with the new methods. was: HADOOP-9252 slightly changes the format of some StringUtils outputs. It may cause test failures. Also, some methods was deprecated by HADOOP-9252. The use of them should be replaced with the new methods. Fix test failure for HADOOP-9252 Key: YARN-377 URL: https://issues.apache.org/jira/browse/YARN-377 Project: Hadoop YARN Issue Type: Bug Reporter: Tsz Wo (Nicholas), SZE Priority: Minor HADOOP-9252 slightly changes the format of some StringUtils outputs. It may cause test failures. Also, some methods were deprecated by HADOOP-9252. The use of them should be replaced with the new methods. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Assigned] (YARN-377) Fix test failure for HADOOP-9252
[ https://issues.apache.org/jira/browse/YARN-377?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ] Chris Nauroth reassigned YARN-377: -- Assignee: Chris Nauroth Fix test failure for HADOOP-9252 Key: YARN-377 URL: https://issues.apache.org/jira/browse/YARN-377 Project: Hadoop YARN Issue Type: Bug Reporter: Tsz Wo (Nicholas), SZE Assignee: Chris Nauroth Priority: Minor HADOOP-9252 slightly changes the format of some StringUtils outputs. It may cause test failures. Also, some methods were deprecated by HADOOP-9252. The use of them should be replaced with the new methods. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (YARN-360) Allow apps to concurrently register tokens for renewal
[ https://issues.apache.org/jira/browse/YARN-360?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13570746#comment-13570746 ] Siddharth Seth commented on YARN-360: - +1. Committing this. Nice unit test btw. The test failure isn't related; passes locally on multiple runs. Allow apps to concurrently register tokens for renewal -- Key: YARN-360 URL: https://issues.apache.org/jira/browse/YARN-360 Project: Hadoop YARN Issue Type: Bug Affects Versions: 0.23.3, 3.0.0, 2.0.0-alpha Reporter: Daryn Sharp Assignee: Daryn Sharp Priority: Critical Attachments: YARN-357.patch, YARN-360.patch {{DelegationTokenRenewer#addApplication}} has an unnecessary {{synchronized}} keyword. This serializes job submissions and can add unnecessary latency and/or hang all submissions if there are problems renewing the token. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (YARN-370) CapacityScheduler app submission fails when min alloc size not multiple of AM size
[ https://issues.apache.org/jira/browse/YARN-370?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13570749#comment-13570749 ] Vinod Kumar Vavilapalli commented on YARN-370: -- That's correct, the validation today is only done in secure mode. Extending it to non-secure mode is pending - MAPREDUCE-2744. Also the scheduler? The scheduler in use doesn't seem to be normalizing requests correctly. CapacityScheduler app submission fails when min alloc size not multiple of AM size -- Key: YARN-370 URL: https://issues.apache.org/jira/browse/YARN-370 Project: Hadoop YARN Issue Type: Bug Components: capacityscheduler Affects Versions: 2.0.3-alpha Reporter: Thomas Graves Assignee: Zhijie Shen Priority: Blocker I was running 2.0.3-SNAPSHOT with the capacity scheduler configured with minimum allocation size 1G. The AM size was set to 1.5G. I didn't specify resource calculator so it was using DefaultResourceCalculator. The am launch failed with the error below: Application application_1359688216672_0001 failed 1 times due to Error launching appattempt_1359688216672_0001_01. Got exception: RemoteTrace: at LocalTrace: org.apache.hadoop.yarn.exceptions.impl.pb.YarnRemoteExceptionPBImpl: RemoteTrace: at LocalTrace: org.apache.hadoop.yarn.exceptions.impl.pb.YarnRemoteExceptionPBImpl: Unauthorized request to start container. Expected resource memory:2048, vCores:1 but found memory:1536, vCores:1 at org.apache.hadoop.yarn.factories.impl.pb.YarnRemoteExceptionFactoryPBImpl.createYarnRemoteException(YarnRemoteExceptionFactoryPBImpl.java:39) at org.apache.hadoop.yarn.ipc.RPCUtil.getRemoteException(RPCUtil.java:47) at org.apache.hadoop.yarn.server.nodemanager.containermanager.ContainerManagerImpl.authorizeRequest(ContainerManagerImpl.java:383) at org.apache.hadoop.yarn.server.nodemanager.containermanager.ContainerManagerImpl.startContainer(ContainerManagerImpl.java:400) at org.apache.hadoop.yarn.api.impl.pb.service.ContainerManagerPBServiceImpl.startContainer(ContainerManagerPBServiceImpl.java:68) at org.apache.hadoop.yarn.proto.ContainerManager$ContainerManagerService$2.callBlockingMethod(ContainerManager.java:83) at org.apache.hadoop.ipc.ProtobufRpcEngine$Server$ProtoBufRpcInvoker.call(ProtobufRpcEngine.java:454) at org.apache.hadoop.ipc.RPC$Server.call(RPC.java:1014) at org.apache.hadoop.ipc.Server$Handler$1.run(Server.java:1735) at org.apache.hadoop.ipc.Server$Handler$1.run(Server.java:1731) at java.security.AccessController.doPrivileged(Native Method) at javax.security.auth.Subject.doAs(Subject.java:415) at org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:1441) at org.apache.hadoop.ipc.Server$Handler.run(Server.java:1729) at sun.reflect.NativeConstructorAccessorImpl.newInstance0(Native Method) at sun.reflect.NativeConstructorAccessorImpl.newInstance(NativeConstructorAccessorImpl.java:57) at sun.reflect.DelegatingConstructorAccessorImpl.newInstance(DelegatingConstructorAccessorImpl.java:45) at java.lang.reflect.Constructor.newInstance(Constructor.java:525) at org.apache.hadoop.ipc.RemoteException.instantiateException(RemoteException.java:90) at org.apache.hadoop.ipc.RemoteException.unwrapRemoteException(RemoteException.java:57) at org.apache.hadoop.yarn.exceptions.impl.pb.YarnRemoteExceptionPBImpl.unwrapAndThrowException(YarnRemoteExceptionPBImpl.java:123) at org.apache.hadoop.yarn.api.impl.pb.client.ContainerManagerPBClientImpl.startContainer(ContainerManagerPBClientImpl.java:109) at org.apache.hadoop.yarn.server.resourcemanager.amlauncher.AMLauncher.launch(AMLauncher.java:111) at org.apache.hadoop.yarn.server.resourcemanager.amlauncher.AMLauncher.run(AMLauncher.java:255) at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1110) at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:603) at java.lang.Thread.run(Thread.java:722) . Failing the application. It looks like the launchcontext for the app didn't have the resources rounded up. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (YARN-373) Allow an AM to reuse the resources allocated to container for a new container
[ https://issues.apache.org/jira/browse/YARN-373?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13570782#comment-13570782 ] Alejandro Abdelnur commented on YARN-373: - Hitesh, Didn't dive into the whole approach yet, first wanted to 'socialize' the idea. Now let me answer with my current thoughts. For this use case I was not thinking about resizing 'inflight' containers, while we could resize easily on CPU, for memory would be quite difficult. The use case is about shortcutting getting resources for a container by reusing the same (or less) resources being freed up by a terminating container in the same node. By doing this you don't have to go to all the way to the scheduler and compete/wait for those resources to become avail. In short, recycling resources the AM already got. The terminating container would still exit, not changing the notion of completion of a container. The container using the recycled resources would be a fresh new container process. (Otherwise we could not shrink in memory). Regarding localized resources, a new resource localization would be done. Allow an AM to reuse the resources allocated to container for a new container - Key: YARN-373 URL: https://issues.apache.org/jira/browse/YARN-373 Project: Hadoop YARN Issue Type: Improvement Components: resourcemanager Affects Versions: 2.0.3-alpha Reporter: Alejandro Abdelnur Assignee: Alejandro Abdelnur When a container completes, instead the corresponding resources being freed up, it should be possible for the AM to reuse the assigned resources for a new container. As part of the reallocation, the AM would notify the RM about partial resources being freed up and the RM would make the necessary corrections in the corresponding node. With this functionality, an AM can ensure it gets a container in the same node where previous containers run. This will allow getting rid of the ShuffleHandler as a service in the NMs and run it as regular container task of the corresponding AM. In this case, the reallocation would reduce the CPU/MEM obtained for the original container to the what is needed for serving the shuffle. Note that in this example the MR AM would only do this reallocation for one of the many tasks that may have run in a particular node (as a single shuffle task could serve all the map outputs from all map tasks run in that node). -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (YARN-360) Allow apps to concurrently register tokens for renewal
[ https://issues.apache.org/jira/browse/YARN-360?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13570797#comment-13570797 ] Hudson commented on YARN-360: - Integrated in Hadoop-trunk-Commit #3320 (See [https://builds.apache.org/job/Hadoop-trunk-Commit/3320/]) YARN-360. Allow apps to concurrently register tokens for renewal. Contributed by Daryn Sharp. (Revision 1442441) Result = SUCCESS sseth : http://svn.apache.org/viewcvs.cgi/?root=Apache-SVNview=revrev=1442441 Files : * /hadoop/common/trunk/hadoop-yarn-project/CHANGES.txt * /hadoop/common/trunk/hadoop-yarn-project/hadoop-yarn/dev-support/findbugs-exclude.xml * /hadoop/common/trunk/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/security/DelegationTokenRenewer.java * /hadoop/common/trunk/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/test/java/org/apache/hadoop/yarn/server/resourcemanager/security/TestDelegationTokenRenewer.java Allow apps to concurrently register tokens for renewal -- Key: YARN-360 URL: https://issues.apache.org/jira/browse/YARN-360 Project: Hadoop YARN Issue Type: Bug Affects Versions: 0.23.3, 3.0.0, 2.0.0-alpha Reporter: Daryn Sharp Assignee: Daryn Sharp Priority: Critical Attachments: YARN-357.patch, YARN-360.patch {{DelegationTokenRenewer#addApplication}} has an unnecessary {{synchronized}} keyword. This serializes job submissions and can add unnecessary latency and/or hang all submissions if there are problems renewing the token. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Updated] (YARN-366) Add a tracing async dispatcher to simplify debugging
[ https://issues.apache.org/jira/browse/YARN-366?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ] Sandy Ryza updated YARN-366: Attachment: YARN-366.patch Add a tracing async dispatcher to simplify debugging Key: YARN-366 URL: https://issues.apache.org/jira/browse/YARN-366 Project: Hadoop YARN Issue Type: New Feature Components: nodemanager, resourcemanager Affects Versions: 2.0.2-alpha Reporter: Sandy Ryza Assignee: Sandy Ryza Attachments: YARN-366.patch Exceptions thrown in YARN/MR code with asynchronous event handling do not contain informative stack traces, as all handle() methods sit directly under the dispatcher thread's loop. This makes errors very difficult to debug for those who are not intimately familiar with the code, as it is difficult to see which chain of events caused a particular outcome. I propose adding an AsyncDispatcher that instruments events with tracing information. Whenever an event is dispatched during the handling of another event, the dispatcher would annotate that event with a pointer to its parent. When the dispatcher catches an exception, it could reconstruct a stack trace of the chain of events that led to it, and be able to log something informative. This would be an experimental feature, off by default, unless extensive testing showed that it did not have a significant performance impact. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (YARN-366) Add a tracing async dispatcher to simplify debugging
[ https://issues.apache.org/jira/browse/YARN-366?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13570856#comment-13570856 ] Sandy Ryza commented on YARN-366: - I've attached an initial patch that adds TracingAsyncDispatcher. The basic idea of it is that it maintains a thread local reference to the event currently being handled, so that any other events that are fired off while it is being handled get tagged with it as a parent. I've added an EventTrace field to the Event class that maintains trace and parentage information - if we want this feature to entirely not affect existing code, we could instead maintain a mapping inside the dispatcher. I still need to add in configuration hooks to turn it on. Add a tracing async dispatcher to simplify debugging Key: YARN-366 URL: https://issues.apache.org/jira/browse/YARN-366 Project: Hadoop YARN Issue Type: New Feature Components: nodemanager, resourcemanager Affects Versions: 2.0.2-alpha Reporter: Sandy Ryza Assignee: Sandy Ryza Attachments: YARN-366.patch Exceptions thrown in YARN/MR code with asynchronous event handling do not contain informative stack traces, as all handle() methods sit directly under the dispatcher thread's loop. This makes errors very difficult to debug for those who are not intimately familiar with the code, as it is difficult to see which chain of events caused a particular outcome. I propose adding an AsyncDispatcher that instruments events with tracing information. Whenever an event is dispatched during the handling of another event, the dispatcher would annotate that event with a pointer to its parent. When the dispatcher catches an exception, it could reconstruct a stack trace of the chain of events that led to it, and be able to log something informative. This would be an experimental feature, off by default, unless extensive testing showed that it did not have a significant performance impact. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (YARN-366) Add a tracing async dispatcher to simplify debugging
[ https://issues.apache.org/jira/browse/YARN-366?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13570865#comment-13570865 ] Hadoop QA commented on YARN-366: {color:green}+1 overall{color}. Here are the results of testing the latest attachment http://issues.apache.org/jira/secure/attachment/12567927/YARN-366.patch against trunk revision . {color:green}+1 @author{color}. The patch does not contain any @author tags. {color:green}+1 tests included{color}. The patch appears to include 1 new or modified test files. {color:green}+1 javac{color}. The applied patch does not increase the total number of javac compiler warnings. {color:green}+1 javadoc{color}. The javadoc tool did not generate any warning messages. {color:green}+1 eclipse:eclipse{color}. The patch built with eclipse:eclipse. {color:green}+1 findbugs{color}. The patch does not introduce any new Findbugs (version 1.3.9) warnings. {color:green}+1 release audit{color}. The applied patch does not increase the total number of release audit warnings. {color:green}+1 core tests{color}. The patch passed unit tests in hadoop-yarn-project/hadoop-yarn/hadoop-yarn-common. {color:green}+1 contrib tests{color}. The patch passed contrib unit tests. Test results: https://builds.apache.org/job/PreCommit-YARN-Build/379//testReport/ Console output: https://builds.apache.org/job/PreCommit-YARN-Build/379//console This message is automatically generated. Add a tracing async dispatcher to simplify debugging Key: YARN-366 URL: https://issues.apache.org/jira/browse/YARN-366 Project: Hadoop YARN Issue Type: New Feature Components: nodemanager, resourcemanager Affects Versions: 2.0.2-alpha Reporter: Sandy Ryza Assignee: Sandy Ryza Attachments: YARN-366.patch Exceptions thrown in YARN/MR code with asynchronous event handling do not contain informative stack traces, as all handle() methods sit directly under the dispatcher thread's loop. This makes errors very difficult to debug for those who are not intimately familiar with the code, as it is difficult to see which chain of events caused a particular outcome. I propose adding an AsyncDispatcher that instruments events with tracing information. Whenever an event is dispatched during the handling of another event, the dispatcher would annotate that event with a pointer to its parent. When the dispatcher catches an exception, it could reconstruct a stack trace of the chain of events that led to it, and be able to log something informative. This would be an experimental feature, off by default, unless extensive testing showed that it did not have a significant performance impact. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira