[CASSANDRA-13666] Secondary index query on partition key columns might not return partitions with only static data

• Type: Bug
• Status: Resolved
• Priority: Normal
• Resolution: Fixed
• Fix Version/s: 3.0.21, 3.11.7, 4.0-beta1, 4.0
• Component/s: Feature/2i Index
• Severity: Normal
• Since Version: 3.0.0

Source Control Link

https://github.com/apache/cassandra/commit/86e1590042116b35a63a705676ecdffd5dfcde6c

Test and Documentation Plan

Attched to PRs

• Assignee: Berenguer Blasi
• Reporter: Benjamin Lerer

Authors

Berenguer Blasi

Reviewers

Benjamin Lerer

• Created: 04/Jul/17 15:07
• Updated: 14/Jul/20 21:50
• Resolved: 20/May/20 09:35

Description

The problem can be reproduced with the following test in 3.0:

   @Test
    public void testIndexOnPartitionKeyWithPartitionWithoutRows() throws Throwable
    {
        createTable("CREATE TABLE %s (pk1 int, pk2 int, c int, s int static, v int, PRIMARY KEY((pk1, pk2), c))");
        createIndex("CREATE INDEX ON %s (pk2)");

        execute("INSERT INTO %s (pk1, pk2, c, s, v) VALUES (?, ?, ?, ?, ?)", 1, 1, 1, 9, 1);
        execute("INSERT INTO %s (pk1, pk2, c, s, v) VALUES (?, ?, ?, ?, ?)", 1, 1, 2, 9, 2);
        execute("INSERT INTO %s (pk1, pk2, c, s, v) VALUES (?, ?, ?, ?, ?)", 3, 1, 1, 9, 1);
        execute("INSERT INTO %s (pk1, pk2, c, s, v) VALUES (?, ?, ?, ?, ?)", 4, 1, 1, 9, 1);
        flush();

        assertRows(execute("SELECT * FROM %s WHERE pk2 = ?", 1),
                   row(1, 1, 1, 9, 1),
                   row(1, 1, 2, 9, 2),
                   row(3, 1, 1, 9, 1),
                   row(4, 1, 1, 9, 1));

        execute("DELETE FROM %s WHERE pk1 = ? AND pk2 = ? AND c = ?", 3, 1, 1);

        assertRows(execute("SELECT * FROM %s WHERE pk2 = ?", 1),
                   row(1, 1, 1, 9, 1),
                   row(1, 1, 2, 9, 2),
                   row(3, 1, null, 9, null),  // This row will not be returned
                   row(4, 1, 1, 9, 1));
    }

The problem seems to be that the index entries for the static data are inserted with an empty clustering key. When the first SELECT is executed those entries are removed by CompositesSearcher::filterStaleEntries which consider that those entries are stales. When the second SELECT is executed the index ignore the (3, 1) partition as there is not entry for it anymore.