zclllyybb commented on issue #65356: URL: https://github.com/apache/doris/issues/65356#issuecomment-4911431648
Breakwater-GitHub-Analysis-Slot: slot_006b7425fa13 This content is generated by AI for reference only. Initial triage for apache/doris#65356: I reread the live issue. It is open, currently has no labels, no comments, and no affected version field, so I inspected current `upstream/master` read-only. Preliminary judgment: this looks like a real analyzer/executor mismatch around `ARRAY_UNION`, not just an unhelpful message. Current public regression coverage already expects `ARRAY_UNION(array_array, array_array)` to be rejected with `array_union does not support types: ARRAY<ARRAY<INT>>`, so nested array elements are not consistently supported by `ARRAY_UNION` today. The empty-left-array shape can bypass that clean rejection because `ArrayUnion.checkLegalityBeforeTypeCoercion()` only checks argument 0 before type coercion. If argument 0 is an empty array literal, it may not expose the final nested-array item type until coercion with the right argument. Relevant current code path: - FE: `fe/fe-core/src/main/java/org/apache/doris/nereids/trees/expressions/functions/scalar/ArrayUnion.java` checks only `getArgument(0)` for complex/variant/json array item types. - BE: `be/src/exprs/function/array/function_array_union.cpp` registers `FunctionArrayNary<ArrayMapImpl<MapOperation::UNION>, NameArrayUnion>`. - BE dispatch: `be/src/exprs/function/array/function_array_map.h` executes only `_execute_internal<ALL_COLUMNS_SIMPLE>`. `ALL_COLUMNS_SIMPLE` is numeric/time/string/IP columns, not `ColumnArray`. When the item type is itself an array, the nested column is `ColumnArray`, so the dispatch cannot match and falls back to `Unexpected columns`. Recommended next step: 1. Decide the intended contract for `ARRAY_UNION` on nested arrays. If nested arrays should be supported, the fix should be in the BE set/map implementation, with real hash/equality/copy handling for `ColumnArray` elements, including nested offsets and nested null maps. 2. If nested arrays should remain unsupported for `ARRAY_UNION`, the minimum fix is to make the analyzer reject all argument positions cleanly, not only argument 0. `ArrayIntersect.checkLegalityBeforeTypeCoercion()` is a closer existing pattern because it iterates over all arguments and handles null/empty-array typing more deliberately. 3. In either direction, the BE should not expose raw `Unexpected columns` for a user-level type/capability miss. Suggested regression coverage: - `array_union([], [[[1], [2]]])` - a non-empty nested-array left input with nested-array right input - varargs with one empty array and one nested-array argument - nullable nested elements, if nested support is chosen - ordering-insensitive verification, since the current union path materializes results from a hash map Missing information to fully close the issue: - Exact failing SQL and the full returned error text. - Affected Doris version or commit. - Whether the failure is Nereids-only. - Expected duplicate-comparison semantics for nested arrays if support is added. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
