[ https://issues.apache.org/jira/browse/CASSANDRA-6825?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13951437#comment-13951437 ]
Tyler Hobbs commented on CASSANDRA-6825: ---------------------------------------- It loos like CASSANDRA-6327 is the cause for this. The logic for testing sstables for inclusion when there's a composite comparator and multiple components in the slice filter is off. This showed up for {{count(\*)}} because counting queries are always paged internally; the second page was erroneously skipping an sstable. If the {{select *}} query has the same page size (10k), it will also omit results. > COUNT(*) with WHERE not finding all the matching rows > ----------------------------------------------------- > > Key: CASSANDRA-6825 > URL: https://issues.apache.org/jira/browse/CASSANDRA-6825 > Project: Cassandra > Issue Type: Bug > Components: Core > Environment: quad core Windows7 x64, single node cluster > Cassandra 2.0.5 > Reporter: Bill Mitchell > Assignee: Tyler Hobbs > Attachments: cassandra.log, selectpartitions.zip, > selectrowcounts.txt, testdb_1395372407904.zip, testdb_1395372407904.zip > > > Investigating another problem, I needed to do COUNT(*) on the several > partitions of a table immediately after a test case ran, and I discovered > that count(*) on the full table and on each of the partitions returned > different counts. > In particular case, SELECT COUNT(*) FROM sr LIMIT 1000000; returned the > expected count from the test 99999 rows. The composite primary key splits > the logical row into six distinct partitions, and when I issue a query asking > for the total across all six partitions, the returned result is only 83999. > Drilling down, I find that SELECT * from sr WHERE s = 5 AND l = 11 AND > partition = 0; returns 30,000 rows, but a SELECT COUNT(*) with the identical > WHERE predicate reports only 14,000. > This is failing immediately after running a single small test, such that > there are only two SSTables, sr-jb-1 and sr-jb-2. Compaction never needed to > run. > In selectrowcounts.txt is a copy of the cqlsh output showing the incorrect > count(*) results. -- This message was sent by Atlassian JIRA (v6.2#6252)