Currently:
- all the PROJECTION nodes have the same cost, which I believe is incorrect. PROJECTION_COVERED should not be doing any extensive computations, it just converts the index entries into documents
- the cost of the projection nodes does not depend on the number of columns projected.
- the cost of PROJECTION is about the same as the cost of the COLLSCAN, which seems incorrect.
Generally, the costs should be such that an IXSCAN + PROJECTION_COVERED should always win on the basis of cost.
- is duplicated by
-
SERVER-108122 Design workload for PROJECTION_* nodes
-
- Closed
-
- is related to
-
-
- Open
-
- related to
-
SERVER-101520 Calibrate the cost of collection scan vs full covering index scan
-
- Closed
-