Alessandro Solimando created CALCITE-7843:
---------------------------------------------
Summary: RelMdExpressionLineage should support Window
Key: CALCITE-7843
URL: https://issues.apache.org/jira/browse/CALCITE-7843
Project: Calcite
Issue Type: Bug
Components: core
Affects Versions: 1.42.0
Reporter: Alessandro Solimando
Assignee: Alessandro Solimando
RelMdExpressionLineage has no handler for {_}Window{_}, so it falls back to the
catch-all _RelNode_ handler
([here|https://github.com/apache/calcite/blob/6484bed4572adbc219b171a9efbbd08a1b662f3a/core/src/main/java/org/apache/calcite/rel/metadata/RelMdExpressionLineage.java#L101-L104]),
which always returns null. This happens even for columns that _Window_ simply
passes through from its input.
_RelMdTableReferences_ already supports Window
([here|https://github.com/apache/calcite/blob/6484bed4572adbc219b171a9efbbd08a1b662f3a/core/src/main/java/org/apache/calcite/rel/metadata/RelMdTableReferences.java#L276-L290]),
so the two metadata types are not consistent. Any consumer of expression
lineage, such as materialized view rewriting, fails on plans containing a
{_}Window{_}.
Reproducer (for {_}RelMetadataTest{_}):
{code:java}
@Test void testExpressionLineageWindow() {
final RelNode rel = sql("select empno, row_number() over (order by sal) from
emp")
.toRel();
final HepProgram program = new HepProgramBuilder()
.addRuleInstance(CoreRules.PROJECT_TO_LOGICAL_PROJECT_AND_WINDOW)
.build();
final HepPlanner planner = new HepPlanner(program);
planner.setRoot(rel);
final RelNode window = planner.findBestExp().getInput(0);
assertThat(window, instanceOf(Window.class));
final RelMetadataQuery mq = window.getCluster().getMetadataQuery();
final RexNode ref = RexInputRef.of(0, window.getRowType().getFieldList());
final Set<RexNode> r = mq.getExpressionLineage(window, ref);
assertThat(String.valueOf(r), is("[[CATALOG, SALES, EMP].#0.$0]"));
}
{code}
The test currently fails as follows:
{noformat}
Expected: is "[[CATALOG, SALES, EMP].#0.$0]"
but: was "null"
{noformat}
Proposed fix: add a _Window_ handler to {_}RelMdExpressionLineage{_}. The
output of a _Window_ is all its input columns, in the same order, followed by
one column for each window function call. If the expression references only
input columns, the handler delegates to the input, the same way the _Sort_
handler does
([here|https://github.com/apache/calcite/blob/6484bed4572adbc219b171a9efbbd08a1b662f3a/core/src/main/java/org/apache/calcite/rel/metadata/RelMdExpressionLineage.java#L422-L425]).
If it references a window function call column, the handler returns null.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
