Details
-
Bug
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
None
-
None
-
None
Description
Couple of issues noticed with checkpointing of large jobs:
1) Task ID of master appears to be important. In most cases it is 0, however sometimes it is not and as we can not control it checkpointing should not depend on it.
2) Race condition happens on master when worker dies:
org.apache.zookeeper.KeeperException$NoNodeException: KeeperErrorCode = NoNode for /_hadoopBsp/job_201411061513.38895_0001/_applicationAttemptsDir/0/_superstepDir/9/_workerHealthyDir/hadoop4921.prn2.facebook.com_3
at org.apache.zookeeper.KeeperException.create(KeeperException.java:111)
at org.apache.zookeeper.KeeperException.create(KeeperException.java:51)
at org.apache.zookeeper.ZooKeeper.getData(ZooKeeper.java:1151)
at org.apache.zookeeper.ZooKeeper.getData(ZooKeeper.java:1180)
at org.apache.giraph.zk.ZooKeeperExt.getData(ZooKeeperExt.java:470)
at org.apache.giraph.utils.WritableUtils.readFieldsFromZnode(WritableUtils.java:126)