Describe the bug
After upgrading from DataFusion 47 to a newer version I started seeing schema mismatch errors caused by updated array type coercion logic that does not preserve nullability information for nested types.
SELECT offset[2]-offset[1] FROM rd;
Arrow error: Invalid argument error: column types must match schema types, expected List(Field { name: "item", data_type: Int32, nullable: true, dict_id: 0, dict_is_ordered: false, metadata: {} }) but found List(Field { name: "item", data_type: Int32, nullable: false, dict_id: 0, dict_is_ordered: false, metadata: {} }) at column index 0
To Reproduce
The following unit test can be used to verify this behavior.
assertion left == right failed
left: [[List(Field { name: "item", data_type: Int64, nullable: true, dict_id: 0, dict_is_ordered: false, metadata: {} }), List(Field { name: "item", data_type: Int64, nullable: true, dict_id: 0, dict_is_ordered: false, metadata: {} })]]
right: [[List(Field { name: "item", data_type: Int64, nullable: false, dict_id: 0, dict_is_ordered: false, metadata: {} }), List(Field { name: "item", data_type: Int64, nullable: false, dict_id: 0, dict_is_ordered: false, metadata: {} })]]
stack backtrace:
fn test_get_valid_types_fixed_size_arrays() -> Result<()> {
let function = "fixed_size_arrays";
let signature = Signature::arrays(2, None, Volatility::Immutable);
let data_types = vec![
DataType::new_fixed_size_list(DataType::Int64, 3, false),
DataType::new_list(DataType::Int32, false),
];
assert_eq!(
get_valid_types(function, &signature.type_signature, &data_types)?,
vec![vec![
DataType::new_list(DataType::Int64, false),
DataType::new_list(DataType::Int64, false),
]]
);
Ok(())
}
This can also be observed by adding additional tracing into coerce_arguments_for_signature_with_scalar_udf. Observe data_type: Int32, nullable: false has changed to data_type: Int32, nullable: true in coerced type.
/// Returns `expressions` coerced to types compatible with
/// `signature`, if possible.
///
/// See the module level documentation for more detail on coercion.
fn coerce_arguments_for_signature_with_scalar_udf(
expressions: Vec<Expr>,
schema: &DFSchema,
func: &ScalarUDF,
) -> Result<Vec<Expr>> {
if expressions.is_empty() {
return Ok(expressions);
}
let current_types = expressions
.iter()
.map(|e| e.get_type(schema))
.collect::<Result<Vec<_>>>()?;
let new_types = data_types_with_scalar_udf(¤t_types, func)?;
println!("schema: {:?}", schema);
println!("current_types: {:?}", current_types);
println!("Coerced types: {:?}", new_types);
expressions
.into_iter()
.enumerate()
.map(|(i, expr)| expr.cast_to(&new_types[i], schema))
.collect()
}
schema: DFSchema { inner: Schema { fields: [Field { name: "offset", data_type: FixedSizeList(Field { name: "item", data_type: Int32, nullable: false, dict_id: 0, dict_is_ordered: false, metadata: {} }, 2), nullable: true, dict_id: 0, dict_is_ordered: false, metadata: {} }], metadata: {"content_computed_columns": "content_embedding,content_offset"} }, field_qualifiers: [Some(Bare { table: "rd" })], functional_dependencies: FunctionalDependencies { deps: [] } }
current_types: [FixedSizeList(Field { name: "item", data_type: Int32, nullable: false, dict_id: 0, dict_is_ordered: false, metadata: {} }, 2), Int64]
Coerced types: [List(Field { name: "item", data_type: Int32, nullable: true, dict_id: 0, dict_is_ordered: false, metadata: {} }), Int64]
Expected behavior
No response
Additional context
The original (correct) behavior was changed by the following improvement:
#15149 (comment)
Describe the bug
After upgrading from DataFusion 47 to a newer version I started seeing schema mismatch errors caused by updated array type coercion logic that does not preserve nullability information for nested types.
To Reproduce
The following unit test can be used to verify this behavior.
This can also be observed by adding additional tracing into
coerce_arguments_for_signature_with_scalar_udf. Observedata_type: Int32, nullable: falsehas changed todata_type: Int32, nullable: truein coerced type.Expected behavior
No response
Additional context
The original (correct) behavior was changed by the following improvement:
#15149 (comment)