Data Type Support

Algeon checks Arrow types before a relational candidate enters native execution. A candidate with an unsupported type keeps the DataFusion baseline, and native_required mode rejects it instead. Operator and function support is on Native SQL Support. The plan for one query is checked with algeon_explain_coverage.

The table uses two levels of support.

Support	Where the column may appear
Full	Anywhere the operator allows: expressions, predicates, join keys, grouping keys, sort keys, aggregate and window inputs.
Carry	Read from a source, selected, aliased, kept through row selection, UNION ALL and the non-key side of a join or aggregation, and returned unchanged.

A candidate that uses a Carry column as an expression value, a key, a sort or partition value, a literal, or an aggregate or window input keeps the DataFusion baseline. A global COUNT(column) without DISTINCT, FILTER or GROUP BY over a List, LargeList or FixedSizeList column is native: it reads only the top-level validity of the column.

Relational columns

Arrow type	Support	Notes
Boolean	Full
Int8, Int16, Int32, Int64, UInt8, UInt16, UInt32, UInt64	Full
Float32, Float64	Full
Decimal128(P, S)	Full	Valid Arrow precision and scale.
Utf8, LargeUtf8, Utf8View	Full
Date32, Date64	Full	Some temporal functions accept Date32 but not Date64. See the function table.
Time32, Time64	Full	At their Arrow resolutions.
Timestamp without a time zone	Full	Second to nanosecond resolution.
Duration	Full	Second to nanosecond resolution.
Timestamp with a time zone	Carry
Decimal32(P, S), Decimal64(P, S)	Carry	Returned as declared when the query has an Arrow schema. Untyped cuDF output maps to Decimal128(38, S).
List<T>, LargeList<T>	Carry	T is Boolean, an integer type, Float32, Float64 or a string type. Output keeps List versus LargeList, the child field and nullability.
FixedSizeList<T, D>	Carry	T is an integer type, Float32 or Float64. D is positive. Output keeps D, the child field and nullability.
Float16, Decimal256	None
Binary, LargeBinary, FixedSizeBinary, BinaryView	None
Interval	None
Struct, Map, Union, Dictionary, ListView, LargeListView, RunEndEncoded	None
List, LargeList, FixedSizeList with any other child	None	No implicit cast of the child type.
Null	None	The untyped null column type.

Native lists use Int32 offsets. Every offset, child length and rows * D must fit the nonnegative Int32 range. A LargeList outside that range is rejected. The range bounds one relation or intermediate result, such as a UDTF relation argument, a join build side or a sort input: at most 2,147,483,647 child values per list column, or 1,048,575 rows of a 2048-wide FixedSizeList. A Parquet scan over a larger file reads it in parts that each fit. A result above the limit fails with size_overflow, naming the column and the row count that fits; filter it, or split it into row ranges combined with UNION ALL. A single Parquet row group above the limit fails planning, and the message gives the row group size to rewrite the file with. Each non-null FixedSizeList row holds exactly D child positions. NaN, infinities, signed zero, root and child nulls, and empty valid lists are payload values that carry paths preserve.

UDTF inputs

Each UDTF family checks its own input types, separately from the relational path. gpu_describe_function('<udtf>') returns the contract and gpu_validate_call('<udtf>', '<call_json>') validates a call before it runs.

cuGraph edge columns

• Source and destination columns use a numeric type listed for the UDTF in the vertex ID matrix, or both are strings (Utf8, LargeUtf8, Utf8View) when the UDTF lists logical string support. String endpoints share one codebook even when their physical types differ, and vertex identity outputs are returned as Utf8.
• Edge weights, when supplied, are Float32 or Float64.
• edge_id_col uses the same numeric domain as the endpoints. String vertex mapping does not support edge_id_col or the include_edge_id* options.

cuVS vector elements

• Vectors arrive as wide feature columns or one list column, and every feature column and vector binding in a call has the same element type. The element and metric matrix lists the supported combinations. Integer IVF-PQ does not support cosine.
• Floating elements must be finite. Elements are not cast implicitly; use CAST in the subquery. Float16 has no cuDF representation.
• Pairwise distance matches its Float32 or Float64 input, ANN distance is Float32 and KMeans cluster_id is Int32.

cuML feature relations

• Features are Float32 or Float64, as wide columns or one list column under any name: a FixedSizeList, or a uniform List or LargeList. Every UDTF except cuml_forest_predict narrows Float64 to Float32 on the device, and the values cuML reads must be finite. cuml_forest_predict takes Float32 only.
• Every relation has an id of type Int32, Int64, Utf8, LargeUtf8 or Utf8View. Repeated IDs are allowed. row_ordinal UInt64 identifies each output row, beside an id of the input type.
• Classifiers take Int32 labels and the regressor finite Float32 targets in training. predict rejects a label column, and unsupervised input treats every non-id column as a feature, including one named label.
• Nullability is not checked in the schema. A null value fails the query with an invalid-input error that names the column.

cuML SQL inputs describes relation bindings and the cuML UDTF index each UDTF's output schema.