[AMBARI-12570] Cluster creates stuck at 9x% (deadlock sql exception)

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Fixed
• Affects Version/s: 2.1.0
• Fix Version/s: 2.1.1
• Assignee: Jonathan Hurley
• Reporter: Jonathan Hurley
• Created: 28/Jul/15 19:47
• Updated: 15/Dec/15 23:37
• Resolved: 05/Aug/15 17:05

Description

Similar to AMBARI-12526, Ambari installation via a blueprint on SQL Azure gets stuck somewhere between 90% and 100% because of a SQL Database deadlock.

This is always between hostcomponentstate.current_state and hostcomponentstate.version.

Rollback reason: 
Local Exception Stack: 
Exception [EclipseLink-4002] (Eclipse Persistence Services - 2.5.2.v20140319-9ad6abd): org.eclipse.persistence.exceptions.DatabaseException
Internal Exception: com.microsoft.sqlserver.jdbc.SQLServerException: Transaction (Process ID 62) was deadlocked on lock resources with another process and has been chosen as the deadlock victim. Rerun the transaction.
Error Code: 1205
Call: UPDATE hostcomponentstate SET current_state = ? WHERE ((((component_name = ?) AND (host_id = ?)) AND (cluster_id = ?)) AND (service_name = ?))
	bind => [5 parameters bound]
	at org.eclipse.persistence.exceptions.DatabaseException.sqlException(DatabaseException.java:331)
	at org.eclipse.persistence.internal.databaseaccess.DatabaseAccessor.executeDirectNoSelect(DatabaseAccessor.java:900)
	at org.eclipse.persistence.internal.databaseaccess.DatabaseAccessor.executeNoSelect(DatabaseAccessor.java:962)
	at org.eclipse.persistence.internal.databaseaccess.DatabaseAccessor.basicExecuteCall(DatabaseAccessor.java:631)
	at org.eclipse.persistence.internal.databaseaccess.ParameterizedSQLBatchWritingMechanism.executeBatch(ParameterizedSQLBatchWritingMechanism.java:149)
	at org.eclipse.persistence.internal.databaseaccess.ParameterizedSQLBatchWritingMechanism.executeBatchedStatements(ParameterizedSQLBatchWritingMechanism.java:134)
	at org.eclipse.persistence.internal.databaseaccess.DatabaseAccessor.writesCompleted(DatabaseAccessor.java:1836)
	at org.eclipse.persistence.internal.sessions.AbstractSession.writesCompleted(AbstractSession.java:4244)
	at org.eclipse.persistence.internal.sessions.UnitOfWorkImpl.writesCompleted(UnitOfWorkImpl.java:5594)
	at org.eclipse.persistence.internal.sessions.RepeatableWriteUnitOfWork.writeChanges(RepeatableWriteUnitOfWork.java:453)
	at org.eclipse.persistence.internal.jpa.EntityManagerImpl.flush(EntityManagerImpl.java:863)
	at org.eclipse.persistence.internal.jpa.QueryImpl.performPreQueryFlush(QueryImpl.java:963)
	at org.eclipse.persistence.internal.jpa.QueryImpl.executeReadQuery(QueryImpl.java:207)
	at org.eclipse.persistence.internal.jpa.QueryImpl.getSingleResult(QueryImpl.java:517)
	at org.eclipse.persistence.internal.jpa.EJBQueryImpl.getSingleResult(EJBQueryImpl.java:400)
	at org.apache.ambari.server.orm.dao.DaoUtils.selectOne(DaoUtils.java:80)
	at org.apache.ambari.server.orm.dao.StackDAO.find(StackDAO.java:93)
	at org.apache.ambari.server.orm.AmbariLocalSessionInterceptor.invoke(AmbariLocalSessionInterceptor.java:53)
	at org.apache.ambari.server.state.svccomphost.ServiceComponentHostImpl.setStackVersion(ServiceComponentHostImpl.java:1058)
	at org.apache.ambari.server.state.svccomphost.ServiceComponentHostImpl$ServiceComponentHostOpStartedTransition.transition(ServiceComponentHostImpl.java:628)
	at org.apache.ambari.server.state.svccomphost.ServiceComponentHostImpl$ServiceComponentHostOpStartedTransition.transition(ServiceComponentHostImpl.java:610)
	at org.apache.ambari.server.state.fsm.StateMachineFactory$SingleInternalArc.doTransition(StateMachineFactory.java:354)
	at org.apache.ambari.server.state.fsm.StateMachineFactory.doTransition(StateMachineFactory.java:294)
	at org.apache.ambari.server.state.fsm.StateMachineFactory.access$300(StateMachineFactory.java:39)
	at org.apache.ambari.server.state.fsm.StateMachineFactory$InternalStateMachine.doTransition(StateMachineFactory.java:440)
	at org.apache.ambari.server.state.svccomphost.ServiceComponentHostImpl.handleEvent(ServiceComponentHostImpl.java:901)
	at org.apache.ambari.server.state.cluster.ClusterImpl.processServiceComponentHostEvents(ClusterImpl.java:2508)
	at org.apache.ambari.server.orm.AmbariJpaLocalTxnInterceptor.invoke(AmbariJpaLocalTxnInterceptor.java:68)
	at org.apache.ambari.server.actionmanager.ActionScheduler.doWork(ActionScheduler.java:343)
	at org.apache.ambari.server.actionmanager.ActionScheduler.run(ActionScheduler.java:195)
	at java.lang.Thread.run(Thread.java:745)

• We have dual X-locks on hostcomponentstate asking for U-locks when updating the CLUSTERED INDEX.
• Both dual X-locks, from different transactions and different processes, are on the same row (technically impossible) - based on the XML execution plan, we can see that the concurrent UPDATE statements are executing on different rows due to their CLUSTERED INDEX predicate.
• In Java, Ambari has locks which prevent concurrent U- or X-locks on the same row
• Only happens on SQL Server

My best suspicion right now is that we have a key hash collision happening on this table. That's why two processes appear to have the same lock even though they are on different rows.

I was able to use a database dump that I took to compare hash values from hostcomponentstate:

SELECT  %%lockres%% as lock_hash, cluster_id, host_id, service_name, component_name
  FROM hostcomponentstate 
  ORDER BY host_id, service_name, %%lockres%%

lock_hash	cluster_id	host_id	service_name	component_name
(0d4a8b0869f5)	2	1	HDFS	SECONDARY_NAMENODE
(0d4a8b0869f5)	2	1	HDFS	HDFS_CLIENT
(99fb0081b824)	2	1	MAPREDUCE2	HISTORYSERVER
(7086998db3dc)	2	1	YARN	APP_TIMELINE_SERVER
(7086998db3dc)	2	1	YARN	RESOURCEMANAGER
(3bef52323322)	2	1	ZOOKEEPER	ZOOKEEPER_SERVER
...

SQL Server is producing lock hashes that collide! It seems like the issue here is that we are using a CLUSTERED INDEX on 4 columns, 3 of which are always the same (cluster, host, service) in many cases. The only variable is the component name. When the hash gets truncated to 6 bytes, we get duplicates.

So, I think this totally aligns with my suspicions as to why this is only a SQL Server problem since other database don't lock like this. It also makes sense that this is the only table this happens on since we are not using a surrogate PK here. I think we have a few options here:

• Add more columns into the CLUSTERED INDEX in hopes we get a more unique hash. The problem is that the other columns are also basically the same.
• Change the CLUSTERED INDEX to an UNCLUSTERED INDEX (since this is our main query criteria) and use a single, unique BIGINT PK as we do for many other tables. I'm just not sure how SQL Server locks on a row when there is a CLUSTERED INDEX which is not part of the predicate.
• Remove the CLUSTERED INDEX entirely (performance would probably tank)
• It's possible we can try to partition this table differently so that lock space is more unique.
• Change the existing CLUSTERED INDEX so that it disallows row and page level locks, forcing all X-locks to be table-level locks. This would, in theory, prevent the deadlock and would not require us to change any data. But it would introduce a bottleneck on the table for anything more than a single read.

Issue Links

• links to: Reviewboard