feat(cubejs): smart-gen — YAML default, shorter names, smarter skips, no auto pre-aggs

[acmeguy]

Summary

Several improvements to smart-gen so it produces noticeably cleaner, more useful Cube models — fewer rollups, shorter field names, more accurate field-skip rules.

Output format

• Default to YAML (.yml). Falls back to JS only when a cube has a FILTER_PARAMS arrow callback we cannot translate. Identity arrows (v) => v are auto-transpiled to Python lambdas (lambda v: v) so YAML works for the nested-lookup-key path too.
• Filename resolution: explicit file_name wins → reuse existing model's extension on re-run → otherwise .yml.

Cube naming from filters

• Filtering by e.g. event = 'Stockout Ended' with no explicit name now derives stockout_ended.yml. Re-running same filter set updates the same file.

Field naming — shortest-unique resolver

Each field carries an ordered candidate list from leaf → fully-qualified. A new resolver picks the shortest non-clashing candidate; basic columns hold their leaf names and longer-candidate fields advance around them. FILTER_PARAMS refs auto-rebind when a lookup-key dim is renamed.

Before	After
props_color	color
commerce_products_id	id
commerce_products_type	type
basic lat + location.lat	lat, location_lat

Value counting fix for Nested(...) parallel arrays

Profiler's arrayColumnSql switched from uniq() (counts whole-array values) to uniqArray() + arrayFilter (counts elements). value_rows now requires at least one meaningful element — string non-empty, number non-null/non-zero, other non-null — so an Array(Nullable(Bool)) of all NULL no longer reads as 100% populated just because the parallel array has length 1. Numeric arrays also emit minArray/maxArray so the all-zero skip rule works for them.

Auto-skip cascade

cubeBuilder.processColumns now skips:

• STRING/UUID with uniqueValues===1 and lc_values[0] empty / whitespace / '0' / 0
• NUMBER with min === max === 0
• BOOLEAN/Int8 with min === max OR uniqueValues === 1 (catches the customer_facing Int8 always-zero case)

Same cascade mirrored into the nested-AJ children path.

No auto pre-aggregations

buildRawCube and the nested-AJ path no longer emit daily_rollup / monthly_rollup. Heuristic rollups bloat CubeStore and surprise users with hidden refresh schedules. User-added pre-aggs in existing models are preserved through merge.

Pre-existing test fixes (was 4 baseline failures)

• package.json test script: --experimental-test-module-mocks so mock.module() works on Node 22.12 (provisionFraiOS suite).
• Legacy ARRAY JOIN path: surface user-supplied alias as a dimension on the flattened cube (collision-safe).
• buildWhereClause test: assert current "no allowlist ⇒ all tables internal" semantics.
• profileTable emitter test: assert current step names (init / initial_profile / profiling).

Test plan

• node --experimental-test-module-mocks --test 'src/**/__tests__/*.test.js' → 511 / 511 passing
• Smoke test in dev: profile a table with nested Nested(...) columns, generate; confirm shorter field names + .yml output
• Smoke test: filter by event = 'X'; confirm derived cube/file name
• Smoke test: nested-array lookup-key cube; confirm lambda v: v in YAML and queries still resolve
• Verify existing .js smart-gen models keep their .js extension on re-run

Companion PR

Frontend mirror of the skip rules in step-2 auto-select: smartdataHQ/client-v2#feat/smart-gen-field-selection

🤖 Generated with Claude Code