Hi All,
I am trying to write a rule to eliminate a potentially redundant
LogicalSort node.
Consider
SELECT I FROM R1 ORDER BY I;
The initial tree is
LogicalSort(sort0=[$0], dir0=[DESC])
LogicalProject(I=[$0])
MyTableScan(table=[[R1]], expr#0..5=[{inputs}], proj#0..5=[{exprs}])
and after all the transformations are applied it becomes
LogicalSort(sort0=[$0], dir0=[DESC])
MyTableScan(table=[[R1]], expr#0..5=[{inputs}], I=[$t0])
The LogicalProject is merged with the MyTableScan which retains all the
original project expressions inside a local RexProgram instance
If MyTableScan already produces rows in an order that matches the ORDER BY
(for example, an underlying table could have an index that guarantees the
ordering if the scan uses that index to retrieve the rows) then LogicalSort
is not required and could be eliminated.
My approach is to write a transformation rule to match the LogicalSort node
and its onMatch method would use a custom RelShuttleImpl class to visit a
child TableScan. The shuttle can compare the collation fields list with the
list of the project expressions to decide whether the sort order
is compatible with the index order and the sort can be eliminated.
Does this approach sound reasonable?
The problem is that I can't figure out how to establish a correspondence
between the sort's collation fields and the scan's columns. Is there a
proper way? If not, what would be a recommended approach?
Any help is greatly appreciated,
Mike
