[ https://issues.apache.org/jira/browse/HBASE-21487?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16713979#comment-16713979 ]
Allan Yang commented on HBASE-21487: ------------------------------------ {quote} May be we can pass the old_table_descriptor also in ModifyTableProcedure. In MODIFY_TABLE_PREPARE step, we can compare the old_table_descriptor with current_table_descriptor, if they are not same then we can throw exception ... {quote} I think you are right, it is better to compare table descriptors directly. > Concurrent modify table ops can lead to unexpected results > ---------------------------------------------------------- > > Key: HBASE-21487 > URL: https://issues.apache.org/jira/browse/HBASE-21487 > Project: HBase > Issue Type: Bug > Affects Versions: 3.0.0, 2.0.0 > Reporter: Syeda Arshiya Tabreen > Priority: Major > > Concurrent modifyTable or add/delete/modify columnFamily leads to incorrect > result. After HBASE-18893, The behavior of add/delete/modify column family > during concurrent operation is changed compare to branch-1.When one client > is adding cf2 and another one cf3 .. In branch-1 final result will be > cf1,cf2,cf3 but now either cf1,cf2 OR cf1,cf3 will be the outcome depending > on which ModifyTableProcedure executed finally.Its because new table > descriptor is constructed before submitting the ModifyTableProcedure in > HMaster class and its not guarded by any lock. > *Steps to reproduce* > 1.Create table 't' with column family 'f1' > 2.Client-1 and Client-2 requests to add column family 'f2' and 'f3' on table > 't' concurrently. > *Expected Result* > Table should have three column families(f1,f2,f3) > *Actual Result* > Table 't' will have column family either (f1,f2) or (f1,f3) -- This message was sent by Atlassian JIRA (v7.6.3#76005)