perf: bound repeated work in analytic queries

[aptend]

What type of PR is this?

• API-change
• BUG
• Improvement
• Documentation
• Feature
• Test and CI
• Code Refactoring

Which issue(s) this PR fixes:

Related to #26768.

What this PR does / why we need it:

Large analytic plans can repeat the same CTE/grouping work, miss safe early pruning, project unique hash matches row by row, and retain a full partition for a small RANK bound. This PR addresses those mechanisms generically:

• restores CTE and grouping-set sharing only with deterministic full-drain proofs, conservative materialization estimates, statement spill budgets, and legacy fallback;
• keeps grouping-set aggregates on the raw input stream, while ignoring inactive rollup keys during group-cardinality estimation;
• exposes safe common keys from mixed DNF predicates and delivers uncorrelated scalar predicates through joins only when the physical runtime-filter topology is provably local;
• batches projection for unique residual-free inner hash joins;
• extends bounded Partition Top-N to exact RANK <= N semantics by retaining boundary ties, with chunked output, accounted state, and an MORPC v67 compatibility gate.

No query ID is encoded in these rules. Statistics participate only in conservative cost/admission decisions, never in semantic legality. Unsupported, uncertain, mixed-version, or over-budget shapes keep the previous plan. The design and counterexample matrix are recorded in the stats-independent analytic rewrite RFC v10 (exact revision).

Measured on the isolated 1T TPC-DS lab with 128 MiB operator spill limits, merge disabled, a 20 GiB no-swap cgroup, and memory/OOM monitoring:

• Q2 completed in 914.102s;
• Q14a completed in 1059.257s; Q14b changed from the 1800s timeout to 994.934s.

The latest revision removes decimal SUM(SUM(v)) prefix reuse. Checked fixed-width addition is not associative in error semantics: a finer partial sequence can overflow even when the raw coarser input succeeds through cancellation. Dynamic grouping-set expansion remains available because each grouping set consumes raw rows in their original order.

Exact-head validation includes the complete pkg/sql/plan test package, go vet ./pkg/sql/plan, a clean build, and an executed public-SQL differential between shared and legacy grouping-set plans. The regression covers Decimal(38,0) near-bound cancellation in the failing row order, NULL aggregate input, runtime-empty input, and four nested rollup prefixes. CI on the updated head is pending.

[qodo-code-review]

Qodo reviews are paused for this user.

Troubleshooting steps vary by plan Learn more →

On a Teams plan?
Reviews resume once this user has a paid seat and their Git account is linked in Qodo.
Link Git account →

Using GitHub Enterprise Server, GitLab Self-Managed, or Bitbucket Data Center?
These require an Enterprise plan - Contact us
Contact us →

[aptend]

Addressed the P1 on exact head 58ec680c6de96fc8db0e07a22f0716973a65bb95.

I removed decimal aggregate-prefix reuse entirely, including its separate 32 GiB admission path. The remaining grouping-set rewrite aggregates each set from raw input, so it does not introduce SUM(SUM(v)) reassociation. RFC v9 records the checked-overflow counterexample and makes raw-input aggregate order part of the contract.

Added two complementary regressions: a planner-shape test that requires decimal SUM to use raw-input dynamic grouping, and an executed public SQL differential against the legacy plan covering the exact Decimal(38,0) cancellation order, NULL input, runtime-empty input, and four rollup prefixes. The shared and legacy executions returned identical results locally; pkg/sql/plan, go vet ./pkg/sql/plan, and a clean build pass. The revised implementation removes state rather than introducing ownership, so the existing shared-source spill/reset contract is unchanged.

[XuPeng-SH]

GPT-6 medium deep review at exact head 58ec680c6de96fc8db0e07a22f0716973a65bb95:

Source review found no verified correctness blocker in the current planner/executor changes. The prior Decimal SUM reassociation issue is closed: finalized-prefix reuse is gone, grouping sets consume raw input, and the added plan-shape plus SQL differential cases cover cancellation, NULL, empty input, and nested prefixes. I also reviewed the CTE/full-drain and scalar-filter proofs, DNF/semi-containment changes, unique hash-join projection, peer-aware RANK Top-N, and v67 compatibility/fallback paths.

P2 — required implementation-acceptance evidence is incomplete. RFC v9 makes these measurements approval gates and requires exact-revision raw artifacts. The PR body reports only three TPC-DS elapsed times; I could not find the fixed TPC-H control results, base-versus-candidate planner p50/p95 time and allocation data, or per-target rows/bytes scanned, peak query memory, and spill bytes. Please attach the exact-revision report (or mark unavailable scale runs as open under the RFC's exception, with its required plan/result/resource artifacts). The exact-head run 34749464335 also still has UT, SCA, coverage, and BVT checks pending, so the mandatory validation set is not terminal yet.

The RFC states that one exact-head APPROVE accepts the design, implementation, and evidence together. I am therefore requesting changes for the missing acceptance evidence, not claiming an unresolved code defect.