一个萝卜一个坑之YARN

问题描述：

NodeManager1 cpu负载飙高，进程还在但是不再向ResourceManager发送心跳，不断重复下文2的动作。心跳停止一段时间后会重连上RM但是cpu仍然很高，再过一段时间心跳再停，一直循环。

NodeManager的日志解析

1.localizing:container开始从hdfs下载resource，hdfs文件的状态从INIT变成DOWNLOADING。

2018-08-25 16:15:38,592 INFO org.apache.hadoop.yarn.server.nodemanager.containermanager.localizer.LocalizedResource：Resource hdfs://mycluster/user/hdfs/.staging/application_1444990016246_29569/libjars/avro-mapred-hadoop2.jar transitioned from INIT to DOWNLOADING

2.container在localizing过程中被stop或者kill，导致hdfs文件状态保持为DOWNLOADING。
non-zero refcount表示当前没有其他container在使用这个资源，说明这个资源将无法删除。

2018-08-25 19:15:38,592 ERROR org.apache.hadoop.yarn.server.nodemanager.containermanager.localizer.LocalResourcesTrackerImpl: Attempt to remove resource: { { hdfs://mycluster/user/hdfs/.staging/application_1444990016246_29569/libjars/avro-mapred-hadoop2.jar, 1448139497492, FILE, null },pending,[],920074451410033,DOWNLOADING} with non-zero refcount

3.任务已经被kill所以报了CancellationException

2018-08-25 19:25:34,592 WARN org.apache.hadoop.yarn.server.nodemanager.containermanager.localizer.ResourceLocalizationService: {...}failed;
java.util.concurrent.CancellationException

4.一段时间后状态从DOWNLOADING转为FAILED，hdfs资源可以删除

2018-08-25 20:15:38,592 INFO org.apache.hadoop.yarn.server.nodemanager.containermanager.localizer.LocalizedResource：Resource hdfs://mycluster/user/hdfs/.staging/application_1444990016246_29569/libjars/avro-mapred-hadoop2.jar(->/data/nm-local-dir/usercache/hadoop/filecache/5432524/avro-mapred-hadoop2.jar) transitioned from DOWNLOADING to FAILED

5.删除本地缓存的文件(可能已损坏)

2018-08-25 19:15:38,592 INFO org.apache.hadoop.yarn.server.nodemanager.containermanager.localizer.LocalResourcesTrackerImpl:Removed /data/nm-local-dir/usercache/hadoop/filecache/5432524/avro-mapred-hadoop2.jar from localized cache

6.请求的资源不在缓存中，将重新请求

2018-08-25 19:15:38,592 INFO org.apache.hadoop.yarn.server.nodemanager.containermanager.localizer.LocalResourcesTrackerImpl:Container container_152345432_4324_3_4324234 sent RELEASE event on a resource request {hdfs://mycluster/user/hdfs/.staging/application_1444990016246_29569/libjars/avro-mapred-hadoop2.jar,,,} not present in cache

原因总结

container被stop，原因可能是与外部组件rpc失败，或者任务被人为kill等等异常。导致hdfs资源异常无法删除而container又会一直尝试去删除

解决办法

1.Low的办法：手动删除hdfs中无法删除的文件(难实现，不知道删那些文件且很多时操作麻烦)

2.高端的办法：异常的位置

public boolean remove(LocalizedResource rem, DeletionService delService) {
// current synchronization guaranteed by crude RLS event for cleanup
   LocalizedResource rsrc = localrsrc.get(rem.getRequest());
   if (null == rsrc) {
     LOG.error("Attempt to remove absent resource: " + rem.getRequest()
         + " from " + getUser());
     return true;
   }
   if (rsrc.getRefCount() > 0
       || ResourceState.DOWNLOADING.equals(rsrc.getState()) || rsrc != rem) {
     // internal error
     LOG.error("Attempt to remove resource: " + rsrc
         + " with non-zero refcount");
     return false;
   } else { // ResourceState is LOCALIZED or INIT
     localrsrc.remove(rem.getRequest());
     if (ResourceState.LOCALIZED.equals(rsrc.getState())) {
       delService.delete(getUser(), getPathToDelete(rsrc.getLocalPath()));
     }
     decrementFileCountForLocalCacheDirectory(rem.getRequest(), rsrc);
     LOG.info("Removed " + rsrc.getLocalPath() + " from localized cache");
     return true;
   }
 }

ResourceState.DOWNLOADING.equals(rsrc.getState())
文件状态为DOWNLOADING则报错，可在源码中删除这个条件。
参考添加补丁：
https://issues.apache.org/jira/browse/YARN-2902
https://issues.apache.org/jira/secure/attachment/12685562/YARN-2902.patch

3.无敌的办法：重启大法。。。重启nodemanager，spark等任务会自动failover，不会影响线上的业务

公众号：大叔据 。

评论不能及时回复可直接加公众号提问或交流，知无不答，谢谢 。