Skip to content

make_array: relax element-type equality to accept inputs differing only in nested-field nullability #22366

@schenksj

Description

@schenksj

Summary

make_array (in datafusion-functions-nested) panics when called with arrays whose element types share the same shape but differ in nested-field nullability. Spark, Postgres, and arrow::compute::concat all accept this and widen nullable to true in the result type. DataFusion's make_array_inner is stricter, which propagates up to any caller that builds array(...) over heterogeneously-produced child expressions.

Repro symptom

Real-world surfacing in apache/datafusion-comet on a Delta Lake CDF write that builds array(struct(id, b, _change_type=lit(\"delete\")), struct(id, b, _change_type=col(...))) — one arm's _change_type is Utf8 non-nullable (from a literal), another is Utf8 nullable:

panicked at arrow-data-58.2.0/src/transform/mod.rs:422:
assertion `left == right` failed: Arrays with inconsistent types passed to MutableArrayData
 left: Struct([Field { name: \"id\", data_type: Int64, nullable: true },
               Field { name: \"b\",  data_type: Int32 },
               Field { name: \"_change_type\", data_type: Utf8 }])
right: Struct([Field { name: \"id\", data_type: Int64, nullable: true },
               Field { name: \"b\",  data_type: Int32 },
               Field { name: \"_change_type\", data_type: Utf8, nullable: true }])

Stack: make_array_innerMutableArrayData::with_capacities.

Proposal

make_array should accept element types that are equal under nullability-widening (recursively, for nested structs/lists/maps). Concretely:

  • Compute the merged element type by walking each child's DataType and OR-ing the nullable flag at every level (this is essentially Field::try_merge minus the type-promotion arm).
  • Cast each child to the merged type before handing to MutableArrayData.
  • Return ArrayType with containsNull = true if any merge raised a nullability flag.

This matches what coerce_types-style coercion does elsewhere in the planner, but applied at execution time when input arrays still disagree (the planner can't always normalize, e.g. when the array is built from disjoint sources like Delta CDF struct literals).

Why this matters

It blocks native execution of any plan that produces struct elements from multiple sources (CDF writes, UNION ALL inside an array(), manually-constructed plans bypassing TypeCoercion). Workaround today: callers must insert explicit casts upstream, or fall back to a non-DataFusion evaluator — both of which lose perf.

Related caller-side mitigation (for context)

Comet just landed a serde-side decline in 4cb9b4dc that falls back to Spark's JVM evaluator when CreateArray's children have different DataTypes. That fix is conservative but loses native execution. Upstreaming the relaxation here would let downstream projects keep native execution and would help any other Arrow-based engine hitting the same shape.

I can put up a PR if the approach lands well.

🤖 Generated with Claude Code

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions