Cassandra一致性日志分析

接上一篇关于Cassandra一致性的讨论，见：

Cassandra一致性问题及客户端解决方案

 

通过对Cassandra源码添加若干日志，今天应该是找到了问题的根本原因。共三台机器10.130.24.90,10.130.24.91,10.130.24.143

问题原因分析：

日志按照请求的先后顺序截取。表格的主键为（empID, deptID）

日志1：

   请求被发送到91机器，要求插入first_name为eran10003的数据，可以看到时间戳为1392953766795000，取到了first_name为eran10003

DEBUG [Thrift:22] 2014-02-21 11:36:06,795 CassandraServer.java (line 1916) execute_cql3_query:[INSERT INTO employees (empID, deptID, first_name, last_name) VALUES (111, 222, 'eran10003', 'landau');] with clever:QUORUM
TRACE [Thrift:22] 2014-02-21 11:36:06,795 QueryProcessor.java (line 97) Process org.apache.cassandra.cql3.statements.UpdateStatement@1342464f @CL.QUORUM
 INFO [Thrift:22] 2014-02-21 11:36:06,795 StorageProxy.java (line 801) destination:/10.130.24.90
 INFO [Thrift:22] 2014-02-21 11:36:06,795 StorageProxy.java (line 822) insert to different server
 INFO [Thrift:22] 2014-02-21 11:36:06,796 StorageProxy.java (line 831) direct writes to local DC
TRACE [Thrift:22] 2014-02-21 11:36:06,796 MessagingService.java (line 615) /10.130.24.91 sending MUTATION to 29474@/10.130.24.90
 INFO [Thrift:22] 2014-02-21 11:36:06,796 StorageProxy.java (line 801) destination:/10.130.24.143
 INFO [Thrift:22] 2014-02-21 11:36:06,796 StorageProxy.java (line 822) insert to different server
 INFO [Thrift:22] 2014-02-21 11:36:06,796 StorageProxy.java (line 831) direct writes to local DC
TRACE [Thrift:22] 2014-02-21 11:36:06,796 MessagingService.java (line 615) /10.130.24.91 sending MUTATION to 29475@/10.130.24.143
 INFO [Thrift:22] 2014-02-21 11:36:06,796 StorageProxy.java (line 801) destination:/10.130.24.91
 INFO [Thrift:22] 2014-02-21 11:36:06,797 StorageProxy.java (line 817) insertLocal
DEBUG [MutationStage:112] 2014-02-21 11:36:06,797 Keyspace.java (line 351) RowMutation:RowMutation(keyspace='contactks', key='0000006f', modifications=[java.nio.HeapByteBuffer[pos=0 lim=10 cap=12]: timestamp:1392953766795000,java.nio.HeapByteBuffer[pos=0 lim=20 cap=38]:eran10003 timestamp:1392953766795000,java.nio.HeapByteBuffer[pos=0 lim=19 cap=36]:landau timestamp:1392953766795000,])
DEBUG [MutationStage:112] 2014-02-21 11:36:06,797 Keyspace.java (line 359) Appending to commitlog
DEBUG [MutationStage:112] 2014-02-21 11:36:06,797 Keyspace.java (line 375) Adding to employees memtable
DEBUG [Thread-3] 2014-02-21 11:36:06,798 MessagingService.java (line 697) Message received from /10.130.24.90, id :29474
DEBUG [RequestResponseStage:312] 2014-02-21 11:36:06,798 MessageDeliveryTask.java (line 61) doVerb:REQUEST_RESPONSE
DEBUG [Thrift:22] 2014-02-21 11:36:06,798 Tracing.java (line 157) request complete 

  

日志2：

    请求被发送到90机器，要求插入first_name为eran10004的数据，可以看到时间戳为1392953753016000 ，取到了first_name仍然为eran10003

DEBUG [Thrift:16] 2014-02-21 11:35:53,016 org.apache.cassandra.thrift.CassandraServer (line 1916) execute_cql3_query:[INSERT INTO employees (empID, deptID, first_name, last_name) VALUES (111, 222, 'eran10004', 'landau');] with clever:QUORUM
TRACE [Thrift:16] 2014-02-21 11:35:53,016 org.apache.cassandra.cql3.QueryProcessor (line 97) Process org.apache.cassandra.cql3.statements.UpdateStatement@6372d3ed @CL.QUORUM
 INFO [Thrift:16] 2014-02-21 11:35:53,016 org.apache.cassandra.service.StorageProxy (line 801) destination:/10.130.24.90
 INFO [Thrift:16] 2014-02-21 11:35:53,016 org.apache.cassandra.service.StorageProxy (line 817) insertLocal
 INFO [Thrift:16] 2014-02-21 11:35:53,017 org.apache.cassandra.service.StorageProxy (line 801) destination:/10.130.24.143
 INFO [Thrift:16] 2014-02-21 11:35:53,017 org.apache.cassandra.service.StorageProxy (line 822) insert to different server
 INFO [Thrift:16] 2014-02-21 11:35:53,017 org.apache.cassandra.service.StorageProxy (line 831) direct writes to local DC
TRACE [Thrift:16] 2014-02-21 11:35:53,017 org.apache.cassandra.net.MessagingService (line 615) /10.130.24.90 sending MUTATION to 28704@/10.130.24.143
 INFO [Thrift:16] 2014-02-21 11:35:53,017 org.apache.cassandra.service.StorageProxy (line 801) destination:/10.130.24.91
 INFO [Thrift:16] 2014-02-21 11:35:53,018 org.apache.cassandra.service.StorageProxy (line 822) insert to different server
 INFO [Thrift:16] 2014-02-21 11:35:53,018 org.apache.cassandra.service.StorageProxy (line 831) direct writes to local DC
TRACE [Thrift:16] 2014-02-21 11:35:53,018 org.apache.cassandra.net.MessagingService (line 615) /10.130.24.90 sending MUTATION to 28705@/10.130.24.91
DEBUG [MutationStage:117] 2014-02-21 11:35:53,017 org.apache.cassandra.db.Keyspace (line 351) RowMutation:RowMutation(keyspace='contactks', key='0000006f', modifications=[java.nio.HeapByteBuffer[pos=0 lim=10 cap=12]: timestamp:1392953753016000,java.nio.HeapByteBuffer[pos=0 lim=20 cap=38]:eran10004 timestamp:1392953753016000,java.nio.HeapByteBuffer[pos=0 lim=19 cap=36]:landau timestamp:1392953753016000,])
DEBUG [MutationStage:117] 2014-02-21 11:35:53,018 org.apache.cassandra.db.Keyspace (line 359) Appending to commitlog
DEBUG [MutationStage:117] 2014-02-21 11:35:53,019 org.apache.cassandra.db.Keyspace (line 375) Adding to employees memtable
DEBUG [Thread-8] 2014-02-21 11:35:53,020 org.apache.cassandra.net.MessagingService (line 697) Message received from /10.130.24.91, id :28705
DEBUG [Thread-10] 2014-02-21 11:35:53,020 org.apache.cassandra.net.MessagingService (line 697) Message received from /10.130.24.143, id :28704
DEBUG [RequestResponseStage:82] 2014-02-21 11:35:53,021 org.apache.cassandra.net.MessageDeliveryTask (line 61) doVerb:REQUEST_RESPONSE
DEBUG [RequestResponseStage:83] 2014-02-21 11:35:53,021 org.apache.cassandra.net.MessageDeliveryTask (line 61) doVerb:REQUEST_RESPONSE
DEBUG [Thrift:16] 2014-02-21 11:35:53,021 org.apache.cassandra.tracing.Tracing (line 157) request complete

 

 观察上述结果，发现新插入的数据：eran10004，没有成功获取，但是请求确实已经发送到了服务器，并且已经被执行。究竟是什么原因导致的呢？熟悉Cassandra数据存储特点的一定都知道，Cassandra的每一列都是分版本存储的，即插入一列新数据，Cassandra都要赋予其一个时间戳作为版本号，如果主键一致，那原来的数据并不删除，而是直接把新数据覆盖在旧数据上，并按时间戳排序；每次查询都会获取时间戳最大的那一个。

结合日志发现eran10003的时间戳为1392953766795000，大于eran10004的时间戳：1392953753016000 。于是查询的时候，Cassandra就认为eran10003是最新的，而eran10004也就获取不到了。

时间戳是服务器本地时间，91的时间大于90的时间所以出现了问题。

 

解决方案：

 1、使用Ntp同步91和90及143的时间。

 2、时间戳有客户端控制而不在服务端，如果是多客户端，客户端之间还要同步时间。

 当然还需要使用Consistency Level： QUORUM。