[yjhjstz]

Changes

• Skip redundant part-prop requests . Hash and nested-loop joins enumerate a DPE and a non-DPE partition propagation request for every other request. With no partition consumers below the join and no propagation required from above, both give every child an empty spec, so the second one optimizes the same child contexts again. Such requests are now rejected in CEngine::FCheckReqdProps before any child is optimized.
• Reuse child contexts of a request . A request of a group expression is optimized once per context of its group, and under contexts that differ only in the required distribution it usually reaches the same child contexts. The child contexts a request reaches are stored on the group expression, keyed by request number, search stage and the parent's requirements apart from distribution. A later request with an equal key reuses them after recomputing each child's required distribution (Ped) and checking that the child context has a plan, is not the context itself, and is the one a lookup of its requirements finds. As when the children are optimized, the plan properties of earlier children are derived again for the recheck, so it does not change the properties cached on their cost contexts. Cost bound checks after each child are replayed. CTE sequences and multi-stage search are left out.
  • optimizer_enable_child_opt_memo (developer GUC, default on)
  • optimizer_verify_child_opt_memo (developer GUC, default off): optimizes the children anyway and logs any difference from the reused contexts

Planning time

Baseline 13 join/aggregation/subquery/DPE queries. "Part-prop only" is HEAD with optimizer_enable_child_opt_memo=off.

	Serial geomean / sum	Parallel geomean / sum
Part-prop only	−0.6% / −0.8%	−8.7% / −12.2%
HEAD	−8.9% / −10.1%	−20.2% / −25.7%

Parallel/serial planning time ratio (geomean): 1.572 (baseline), 1.445 (part-prop only), 1.377 (HEAD).

Query	Serial, ms	Δ	Parallel, ms	Δ
star4	71.7 → 59.3	−17.2%	137.8 → 93.0	−32.5%
star6	122.3 → 108.4	−11.4%	239.6 → 166.0	−30.7%
star8	223.2 → 208.5	−6.6%	365.3 → 279.7	−23.4%
left3	37.3 → 35.4	−5.3%	49.9 → 42.1	−15.6%
semi	55.5 → 51.3	−7.6%	90.7 → 73.0	−19.6%
anti	29.9 → 29.0	−2.8%	39.6 → 35.8	−9.8%
agg	83.9 → 69.4	−17.4%	184.9 → 120.7	−34.7%
twokey	83.1 → 71.3	−14.1%	176.7 → 121.6	−31.2%
union	14.8 → 14.5	−2.0%	16.9 → 16.1	−4.5%
dpe	11.7 → 11.3	−3.6%	12.7 → 12.2	−3.8%
dpe3	39.1 → 33.7	−13.7%	55.3 → 45.6	−17.5%
corr1	22.6 → 21.4	−5.3%	33.3 → 28.4	−14.5%
corr2	52.8 → 49.0	−7.2%	87.4 → 72.8	−16.6%

What does this PR do?

Type of Change

• Bug fix (non-breaking change)
• New feature (non-breaking change)
• Breaking change (fix or feature with breaking changes)
• Documentation update

Breaking Changes

Test Plan

• Unit tests added/updated
• Integration tests added/updated
• Passed make installcheck
• Passed make -C src/test installcheck-cbdb-parallel

Impact

Performance:

User-facing changes:

Dependencies:

Checklist

• Followed contribution guide
• Added/updated documentation
• Reviewed code for security implications
• This PR contains AI-assisted code generation
• Requested review from cloudberry committers

Additional Context

CI Skip Instructions

---

[my-ship-it]

Great work!

This memoization for child contexts and skipping redundant part-prop requests is a good optimization, and the benchmark results look really promising.

[Alena0704]

Thank you for your work on this! Looks good to me. Great job!