[
https://issues.apache.org/jira/browse/PHOENIX-4841?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16654403#comment-16654403
]
Daniel Wong commented on PHOENIX-4841:
--------------------------------------
Thanks for looking at this [~lhofhansl]. On your two points.
1. This is by design as INVERT should not affect the logical comparison.
2. These nested casts are a by product of two separate processes.
a) During RVC Comparison phoenix checks that both sides of the RVC for each
position in the RVC are castable to a single comparable type. In the case you
pointed out the constant row needed to be casted to DESC to be comparable.
b) The changes in my patch then go through all the values in the row value
constructor and force cast them to ASC order for the comparison.
I decided that handling this case was an optimization that could be handled in
a general manner later.
I know that other databases as an optimization step can detect when the nested
CASTS can be replaced with a single cast. Note that this cannot always be done
depending on if cast allows data modifications. Consider the case of
CAST(CAST('1a' as CHAR(1)) as INTEGER).
> Filters that uses RVC with pk columns where with DESC sort order don't work
> correctly
> -------------------------------------------------------------------------------------
>
> Key: PHOENIX-4841
> URL: https://issues.apache.org/jira/browse/PHOENIX-4841
> Project: Phoenix
> Issue Type: Bug
> Affects Versions: 4.15.0, 5.1.0
> Reporter: Thomas D'Silva
> Assignee: Daniel Wong
> Priority: Major
> Labels: DESC
> Attachments: PHOENIX-4841-v2.patch, PHOENIX-4841-v3.patch,
> PHOENIX-4841-v4.patch, PHOENIX-4841.patch
>
>
> If we filter on pk columns where one of the columns is DESC, we don't get the
> expected results. If the PK columns are of sorted by ASC we get the correct
> results. For eg. the following test fails:
> {code}
> @Test
> public void testRVCWithDescAndAscPK() throws Exception {
> String fullTableName = generateUniqueName();
> // create base table and global view using global connection
> try (Connection conn = DriverManager.getConnection(getUrl())) {
> Statement stmt = conn.createStatement();
> stmt.execute("CREATE TABLE " + fullTableName + "(\n" +
> " A VARCHAR NOT NULL,\n" +
> " B VARCHAR NOT NULL,\n" +
> " C VARCHAR NOT NULL,\n" +
> " CONSTRAINT PK PRIMARY KEY (A, B DESC, C))");
>
> conn.createStatement().execute("UPSERT INTO " + fullTableName + "
> VALUES ('x', 'd', '1')");
> conn.createStatement().execute("UPSERT INTO " + fullTableName + "
> VALUES ('x', 'c', '2')");
> conn.createStatement().execute("UPSERT INTO " + fullTableName + "
> VALUES ('x', 'b', '3')");
> conn.createStatement().execute("UPSERT INTO " + fullTableName + "
> VALUES ('x', 'b', '4')");
> conn.createStatement().execute("UPSERT INTO " + fullTableName + "
> VALUES ('x', 'a', '4')");
> conn.commit();
> }
> // validate that running query using global view gives same results
> try (Connection conn = DriverManager.getConnection(getUrl())) {
> ResultSet rs =
> conn.createStatement().executeQuery(
> "SELECT B, C FROM " + fullTableName + " WHERE (B, C)
> > ('b', '3')");
> assertTrue(rs.next());
> assertEquals("d", rs.getString(1));
> assertEquals("1", rs.getString(2));
> assertTrue(rs.next());
> assertEquals("c", rs.getString(1));
> assertEquals("2", rs.getString(2));
> assertTrue(rs.next());
> assertEquals("b", rs.getString(1));
> assertEquals("4", rs.getString(2));
> assertFalse(rs.next());
> }
> }
> {code}
> The comparison expression for the above query is
> {code}
> (PK[-1], PK[-1]) > (TO_VARCHAR('b'), '3')
> {code}
> When the first row is evaluated the lhs bytes is:
> {code}
> [-101, -1, 49]
> {code}
> and rhs bytes:
> {code}
> [-99, -1, 51]
> {code}
> We invert the bytes of the B column but since the greater than comparison
> operator usedthe row is filtered out (even though it should be returned).
> [~jamestaylor]
> When a column is DESC order do we need to rewrite the comparison expression?
> Instead of
> {code}
> WHERE (B, C) > ('b', '3')
> {code}
> we need something like
> {code}
> WHERE B<~'b' OR (B=~'b' AND C>'3')
> {code}
> Is there a better way to handle this?
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
