[ https://issues.apache.org/jira/browse/CALCITE-4617?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17354350#comment-17354350 ]
Ruben Q L commented on CALCITE-4617: ------------------------------------ [~sylvaincrozon] I have the impression that your original patch was in the right track, we just required more unit tests to verify explicitly that some missing scenarios behave correctly. [~julianhyde], we need to remember that {{SortJoinTransposeRule}} has a limited applicability, and there are a few conditions that must be satisfied in order for the rule to be applied (because otherwise it might produce wrong results): - If join is not a left or right outer, we bail out - If sort is not a trivial order-by, and if there is any sort column that is not part of the input where the sort is pushed, we bail out - If sort has an offset, and if the non-preserved side of the join is not count-preserving against the join condition, we bail out - If the input is already sorted and we are not reducing the number of tuples, we bail out > Wrong offset when SortJoinTransposeRule pushes a sort node with an offset > ------------------------------------------------------------------------- > > Key: CALCITE-4617 > URL: https://issues.apache.org/jira/browse/CALCITE-4617 > Project: Calcite > Issue Type: Bug > Components: core > Affects Versions: 1.26.0 > Reporter: Sylvain Crozon > Assignee: Ruben Q L > Priority: Major > Labels: pull-request-available > Fix For: 1.27.0 > > Time Spent: 1h 50m > Remaining Estimate: 0h > > The SortJoinTransposeRule will push a sort node past a join, and then > duplicate the sort node on top of the join. When the sort node has an offset, > we should only apply it once, otherwise we end up skipping twice as many rows > as we should. The sort node added on top of the join should have a null > offset. > > For example the testSortJoinTranspose6 test checks that for this initial plan > {code} > LogicalProject(DEPTNO=[$0], EMPNO=[$2]) > LogicalSort(offset=[2], fetch=[10]) > LogicalJoin(condition=[=($0, $9)], joinType=[right]) > LogicalTableScan(table=[[CATALOG, SALES, DEPT]]) > LogicalTableScan(table=[[CATALOG, SALES, EMP]]) > {code} > the SortJoinTransposeRule should convert to > {code} > LogicalProject(DEPTNO=[$0], EMPNO=[$2]) > LogicalSort(offset=[2], fetch=[10]) > LogicalJoin(condition=[=($0, $9)], joinType=[right]) > LogicalTableScan(table=[[CATALOG, SALES, DEPT]]) > LogicalSort(offset=[2], fetch=[10]) > LogicalTableScan(table=[[CATALOG, SALES, EMP]]) > {code} > Which will result in applying the offset twice. Instead the LogicalSort on > top of the join should just have a null offset -- This message was sent by Atlassian Jira (v8.3.4#803005)