[Epic] Finer-grained LIR #25989
Labels
A-compute
Area: compute
A-optimization
Area: query optimization and transformation
Epic
[Auto] Multi-stage issue
At the moment, our physical plans contains wide operators that encode several operations in a single instruction. This is somewhat caused by the interface between instruction, which are always well-formed collections or arrangements, but there is little flexibility to break from the pattern.
For example, a delta join requires its inputs to be arranged, and fans out to produce join results based on all inputs. A different approach could be that the individual paths are encoded in LIR, relieving the implementation to making decisions itself. Additionally, it would enable physical optimizations where we could prune some paths that we statically know need not to be arranged.
Product outcome
While not directly user-facing, it would enable future optimizations in the rendering layer that we cannot handle at the moment.
Discovery
Work items
Tasks
Decision log
The text was updated successfully, but these errors were encountered: