fix(tesseract): reduce_by/grain matches time dimensions regardless of granularity

[waralexrom]

Summary

A multi-stage reduce_by / grain reference to a time dimension silently did nothing, producing wrong results with no error. Fixes #10854.

The projected time dimension carries a granularity suffix in its full name (e.g. created_at_month), while the reduce_by reference points at the bare dimension (created_at). Matching was done by full_name string equality, so the two never matched and the time dimension was never removed from the window's PARTITION BY. Every partition collapsed to a single row — a rank measure returned 1 for every row, and an aggregate silently fell back to a plain group-by returning the un-reduced value. String dimensions have no suffix, so they matched and worked; only time dimensions were affected.

Changes

• Add MemberSymbol::matches_grain_reference — behaves like has_member_in_reference_chain, but a time dimension additionally matches through its underlying base dimension, so a bare grain reference removes it regardless of the granularity it is queried at.
• Use it in both grain reshape sites (member_partition_by_logical and partition_filter) for exclude / keep_only.

Testing

• New Rust integration tests (real Postgres) covering the cube and view paths: multi_stage/reduce_by.rs, rank.rs, views.rs. Each asserts a window is used, created_at is excluded from PARTITION BY, and snapshots the executed results (per-group totals / correct ranks). All were red before the fix, green after.
• Full cubesqlplanner suite: 1073 passed, 0 failed.

[claude]

Claude finished @waralexrom's task in 1m 16s —— View job

---

Code Review

LGTM — no blocking issues found (0 high, 0 medium, 1 low)

• Read PR diff and analyze fix
• Verify matches_grain_reference semantics
• Check both call sites are updated symmetrically
• Review test coverage

Summary

The root-cause analysis in the PR description matches the code: projected time dimensions carry a granularity suffix (created_at_month), while reduce_by / grain references point at the bare dimension (created_at), so the old full_name-equality chain check silently missed them. The fix — a dedicated matches_grain_reference that additionally walks through TimeDimension::base_symbol() — is minimal, well-scoped, and correctly applied at both grain-reshape sites (member_partition_by_logical and partition_filter, exclude + keep_only).

Code quality is good: new method is documented with the why, the helper name is clearer than reusing has_member_in_reference_chain, and the change is additive (does not affect string-dimension matching, which already worked).

Correctness

rust/cube/cubesqlplanner/cubesqlplanner/src/planner/symbols/member_symbol.rs:246 — the implementation delegates to has_member_in_reference_chain first, and only falls back to base_symbol() for TimeDimension. base_symbol() returns the underlying dimension symbol; walking its own reference chain covers view aliases and cross-cube joins the same way the existing path does. No behavior change for Dimension / Measure / MemberExpression.

Both call sites (member_query_planner.rs:545,553 and multi_stage_query_planner.rs:297,305) swap the method consistently — exclude and keep_only are updated symmetrically, so there is no asymmetric filter risk.

Testing

test_reduce_by_time_dimension, test_rank_reduce_by_time_dimension, and the corresponding view test each assert three orthogonal properties (window is used, created_at is absent from every PARTITION BY, snapshot of the actual per-partition rows). Snapshots for all three are present. The partition_by_clauses helper is test-only and safe on the generated SQL (no comments/strings containing PARTITION BY in the fixture output).

Low — nit only

partition_by_clauses scans between PARTITION BY and the next ORDER BY (rust/cube/cubesqlplanner/cubesqlplanner/src/tests/integration/multi_stage/reduce_by.rs:32). Fine for these fixtures, but a partition-only window (OVER (PARTITION BY x) with no ORDER BY) or a query with ORDER BY inside a subquery preceding the window would confuse the parser. Not worth changing here — just a note if this helper gets reused. Fix this →

· Branch: `tesseract-reduce_by-fixes`

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 79.26%. Comparing base (5b96418) to head (dc51a64).

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #11188      +/-   ##
==========================================
- Coverage   83.70%   79.26%   -4.45%     
==========================================
  Files         257      474     +217     
  Lines       79465    97013   +17548     
  Branches        0     3572    +3572     
==========================================
+ Hits        66517    76894   +10377     
- Misses      12948    19603    +6655     
- Partials        0      516     +516     

Flag	Coverage Δ
cube-backend	59.13% <ø> (?)
cubesql	83.70% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[claude]

Claude finished @waralexrom's task in 35s —— View job

---

Code Review

LGTM — no blocking issues (0 high, 0 medium, 1 low nit). No new commits since prior review.

• Check for new commits since last review
• Re-verify matches_grain_reference semantics
• Confirm both call sites updated symmetrically
• Review test coverage

Summary

Branch still contains a single commit (dc51a64) — the same diff reviewed on 2026-07-01. Re-verified against the current tree; findings unchanged and the PR already has an approving human review from @ovr.

The fix — a dedicated matches_grain_reference that additionally walks through TimeDimension::base_symbol() — is minimal, well-scoped, and correctly applied at both grain-reshape sites (member_partition_by_logical and partition_filter, exclude + keep_only). No behavior change for Dimension / Measure / MemberExpression.

Correctness

rust/cube/cubesqlplanner/cubesqlplanner/src/planner/symbols/member_symbol.rs:246 — delegates to has_member_in_reference_chain first, falls back to base_symbol() for TimeDimension. Walking the base symbol's own reference chain covers view aliases and cross-cube joins the same way the existing path does.

Both call sites (member_query_planner.rs:545,553 and multi_stage_query_planner.rs:297,305) swap the method consistently — exclude and keep_only are updated symmetrically, so no asymmetric filter risk.

Testing

test_reduce_by_time_dimension, test_rank_reduce_by_time_dimension, and the view test each assert three orthogonal properties (window is used, created_at is absent from every PARTITION BY, snapshot of the per-partition rows). All three snapshots are present.

Low — nit only

partition_by_clauses scans between PARTITION BY and the next ORDER BY (rust/cube/cubesqlplanner/cubesqlplanner/src/tests/integration/multi_stage/reduce_by.rs:32). Fine for these fixtures, but a partition-only window (OVER (PARTITION BY x) with no ORDER BY) or an ORDER BY inside a preceding subquery would confuse the parser. Not worth changing here — just a note if this helper gets reused.

· Branch: tesseract-reduce_by-fixes ·