我使用:Cloudera的經理嵌入式PostgreSQL的蜂巢Metastore服務器的OutOfMemoryError問題
Cloudera Manager Free Edition: 4.5.1
Cloudera Hadoop Distro: CDH 4.2.0-1.cdh4.2.0.p0.10 (Parcel)
Hive Metastore with cloudera manager embedded PostgreSQL database.
我的Cloudera管理器一個單獨的機器上運行,它不是羣集的一部分。
設置使用了Cloudera管理器集羣之後,我開始通過色調+蜂蠟使用蜂巢。
一切都運行良好了一會兒,然後全部suddden的,每當我跑針對有大量的分區(約14000)的特定表中的任何查詢,查詢開始超時:
FAILED: SemanticException org.apache.thrift.transport.TTransportException: java.net.SocketTimeoutException: Read timed out
當我注意到這一點,我看了看日誌,並發現蜂巢Metastore連接已超時:
WARN metastore.RetryingMetaStoreClient: MetaStoreClient lost connection. Attempting to reconnect. org.apache.thrift.transport.TTransportException: java.net.SocketTimeoutException: Read timed out
已經看到了這一點,我想有一個與蜂巢metastore一個問題。所以,我看着日誌蜂巢metastore,發現java.lang.OutOfMemoryErrors:
/var/log/hive/hadoop-cmf-hive1-HIVEMETASTORE-hci-cdh01.hcinsight.net.log.out:
2013-05-07 14:13:08,744 ERROR org.apache.thrift.ProcessFunction: Internal error processing get_partitions_ with_auth
java.lang.OutOfMemoryError: Java heap space
at sun.reflectH.NativeConstructorAccessorImpl.newInstance0(Native Method)
at sun.reflect.NativeConstructorAccessorImpl.newInstance(NativeConstructorAccessorImpl.java:57)
at sun.reflect.DelegatingConstructorAccessorImpl.newInstance(DelegatingConstructorAccessorImpl.jav a:45)
at java.lang.reflect.Constructor.newInstance(Constructor.java:525)
at org.datanucleus.util.ClassUtils.newInstance(ClassUtils.java:95)
at org.datanucleus.store.rdbms.sql.expression.SQLExpressionFactory.newLiteralParameter(SQLExpressi onFactory.java:248)
at org.datanucleus.store.rdbms.scostore.RDBMSMapEntrySetStore.getSQLStatementForIterator(RDBMSMapE ntrySetStore.java:323)
at org.datanucleus.store.rdbms.scostore.RDBMSMapEntrySetStore.iterator(RDBMSMapEntrySetStore.java: 221)
at org.datanucleus.sco.SCOUtils.populateMapDelegateWithStoreData(SCOUtils.java:987)
at org.datanucleus.sco.backed.Map.loadFromStore(Map.java:258)
at org.datanucleus.sco.backed.Map.keySet(Map.java:509)
at org.datanucleus.store.fieldmanager.LoadFieldManager.internalFetchObjectField(LoadFieldManager.j ava:118)
at org.datanucleus.store.fieldmanager.AbstractFetchFieldManager.fetchObjectField(AbstractFetchFiel dManager.java:114)
at org.datanucleus.state.AbstractStateManager.replacingObjectField(AbstractStateManager.java:1183)
at org.apache.hadoop.hive.metastore.model.MStorageDescriptor.jdoReplaceField(MStorageDescriptor.ja va)
at org.apache.hadoop.hive.metastore.model.MStorageDescriptor.jdoReplaceFields(MStorageDescriptor.j ava)
at org.datanucleus.jdo.state.JDOStateManagerImpl.replaceFields(JDOStateManagerImpl.java:2860)
at org.datanucleus.jdo.state.JDOStateManagerImpl.replaceFields(JDOStateManagerImpl.java:2879)
at org.datanucleus.jdo.state.JDOStateManagerImpl.loadFieldsInFetchPlan(JDOStateManagerImpl.java:16 47)
at org.datanucleus.store.fieldmanager.LoadFieldManager.processPersistable(LoadFieldManager.java:63 )
at org.datanucleus.store.fieldmanager.LoadFieldManager.internalFetchObjectField(LoadFieldManager.j ava:84)
at org.datanucleus.store.fieldmanager.AbstractFetchFieldManager.fetchObjectField(AbstractFetchFiel dManager.java:104)
at org.datanucleus.state.AbstractStateManager.replacingObjectField(AbstractStateManager.java:1183)
at org.apache.hadoop.hive.metastore.model.MPartition.jdoReplaceField(MPartition.java)
at org.apache.hadoop.hive.metastore.model.MPartition.jdoReplaceFields(MPartition.java)
at org.datanucleus.jdo.state.JDOStateManagerImpl.replaceFields(JDOStateManagerImpl.java:2860)
at org.datanucleus.jdo.state.JDOStateManagerImpl.replaceFields(JDOStateManagerImpl.java:2879)
at org.datanucleus.jdo.state.JDOStateManagerImpl.loadFieldsInFetchPlan(JDOStateManagerImpl.java:16 47)
at org.datanucleus.ObjectManagerImpl.performDetachAllOnTxnEndPreparation(ObjectManagerImpl.java:35 52)
at org.datanucleus.ObjectManagerImpl.preCommit(ObjectManagerImpl.java:3291)
at org.datanucleus.TransactionImpl.internalPreCommit(TransactionImpl.java:369)
at org.datanucleus.TransactionImpl.commit(TransactionImpl.java:256)
在這一點上,蜂巢metastore得到關閉並重新啓動:
2013-05-07 14:39:40,576 INFO org.apache.hadoop.hive.metastore.HiveMetaStore: Shutting down hive metastore.
2013-05-07 14:41:09,979 INFO org.apache.hadoop.hive.metastore.HiveMetaStore: Starting hive metastore on po rt 9083
現在,爲了解決這個問題,我已經改變了兩個蜂巢metastore服務器和蜂蠟服務器的最大堆大小:
1. Hive/Hive Metastore Server(Base)/Resource Management/Java Heap Size of Metastore Server : 2 GiB (First thing I did.)
2. Hue/Beeswax Server(Base)/Resource Management/Java Heap Size of Beeswax Server : 2 GiB (After reading some groups posts and stuff online, I tried this as well.)
這些都不超過2個步驟似乎已經幫助我繼續看OOMEs中遇到的蜂巢astore日誌。
然後我注意到,實際metastore「數據庫」正在運行作爲我的Cloudera管理器的一部分,如果是PostgreSQL的進程運行內存我不知道。我尋找方法來增加該進程的java堆大小,並發現很少有關於此的文檔。
我在想,如果你們的人能夠幫我解決這個問題。
我應加大對嵌入式數據庫Java堆大小?如果是這樣,我會在哪裏做到這一點?
有沒有別的東西,我失蹤?
謝謝!
忘了提: 當我有很多分區的訪問表此問題只發生。對其他表的查詢運行良好。 – 2013-05-07 23:37:28
此外,當我從配置單元命令行shell運行時也會發生同樣的問題。所以,我猜這與蜜蜂或色相界面無關。 – 2013-05-08 15:58:36
另一個有趣的事情是,當我今天早上將蜂巢Metastore java堆大小增加到4GB並嘗試時,我沒有看到任何OOME(至少現在),但查詢仍然超時。 – 2013-05-08 16:04:53