TJ Banghart created CALCITE-7254:
------------------------------------
Summary: Add a rule for sharing trivially equivelent RelNodes
within Combine
Key: CALCITE-7254
URL: https://issues.apache.org/jira/browse/CALCITE-7254
Project: Calcite
Issue Type: New Feature
Affects Versions: 1.40.0
Reporter: TJ Banghart
Assignee: TJ Banghart
Introduce an optimization rule to identify and consolidate trivially equivalent
{{RelNodes}} that appear within a {{Combine}} operator. Trivial equivalence is
defined as two {{RelNodes}} having identical digests. When such equivalence is
detected, the common subexpression should be lifted into a shared {{Spool}}
reference, and each original occurrence should be replaced with a dependent on
that {{{}Spool{}}}.
An implementation may be able to build on the existing
{{{}RelCommonExpressionSuggester{}}}, which currently identifies common
expressions within individual queries. It may need to be extended to consider
multiple queries or multiple branches within a single {{{}Combine{}}}.
An important consideration will be how {{Combine}} handles cost estimation in
the absence of optimization and how shared expressions might influence that
model. We will likely need to evaluate how these lifted expressions interact
with the cost model to ensure we aren’t over-rewarding shared subtrees or
unintentionally skewing plan selection.
We'd have to take care when managing interactions with other optimization
rules. Filter pushdown, projection trimming, join reordering, etc... can alter
digests and preclude trivial equivalence. The new rule will need safeguards to
prevent conflicts or premature firing relative to these other optimizations.
Some of the implementation details are likely to evolve as this develops.
Initial focus would be on correctness and identifying the right insertion
points in the optimization pipeline.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
