Six follow-up issues from the second-round review of `fused-correctness-review`. All four blockers fixed, both smells fixed, plus regression coverage for the two correctness blockers. Blocker #1 — fallback silently dropped the filter predicate exec_filtered_group_fallback was building `ray_const_table → ray_filter` then overwriting root with `ray_group(...)`, but ray_group has arity 0 and consumes `g->table` directly — the filter node was orphaned. Worse, naively chaining ray_execute(filter) then ray_execute(group) doesn't preserve the selection either: the outer ray_execute compacts and clears g->selection on return, so the group call sees the unfiltered g->table. Fix: take the materialised filtered table from the first ray_execute, swap it in as g->table for the duration of the group call, restore afterwards. Blocker #2 — count1 path accepted nullable group keys Planner gate already rejected this in query.c, but a direct ray_filtered_group() C-API call bypassed the planner. Mirror mk_compile's HAS_NULLS check at exec_filtered_group_count1. Blocker #3 — top-K materialiser corrupted unsupported output types ray_fused_topk_select accepted any column type but the materialise loop reads via read_by_esz / writes via write_col_i64, which only handle scalar BOOL/U8/I16/I32/I64/DATE/TIME/TIMESTAMP/SYM. RAY_STR / GUID / list-shape would gather the raw int64 of the heap-string pointer and write a single byte, producing garbage. Fix: gate output column types at both the planner gate (query.c) and the executor entry (ray_fused_topk_select). Unsupported types fall back to the unfused FILTER + SORT + TAKE. Blocker #4 — top-K sort-key vector parsing ignored SYM width Casting `ray_data(v)` to `int64_t*` is only valid for W64. Compact W8/W16/W32 SYM vectors would read garbage. Use ray_read_sym for the width-correct load. Smell #5 — STR LIKE used the SYM threshold Earlier centralisation accidentally swapped the names; STR path used LIKE_PAR_MIN_ROWS_SYM and SYM path used LIKE_PAR_MIN_ROWS_STR. Restore the correct mapping. Smell #6 — mixed-temporal test was actually same-typed The parity test claimed to cover DATE-vs-TIMESTAMP but used DATE-vs-DATE. Replaced with same-typed DATE compares (which the fused gate accepts) and a comment explaining why a real cross- temporal compare exercises an unsupported engine path that's outside the scope of this branch. Regression coverage: test_fused_group.c - fused_group/fallback_filter_honored: builds a fused node with a nullable agg input (forces fallback) under a selective WHERE; asserts the SUM equals the filtered value, not the unfiltered. - fused_group/count1_rejects_nullable_key: builds a fused node with a nullable group key; verifies the result matches the null-aware unfused group (3 buckets, not 4 with a phantom). fused_group_parity.rfl - Top-K STR output column lex-order check (forces unfused fallback path, asserts correctness). - Same-typed DATE compares verify the fused predicate gate still fires for the supported shape. Tests: 2345 / 2347 passing (1 skipped, 1 pre-existing failure). Bench cluster ratios within noise of the prior state.

Converting a string to a symbol may add random garbage #3

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions