Missing field metadata for projections during physical planning

[linhr]

Describe the bug

DataFusion seems to drop field metadata for projections during physical planning. This may cause schema mismatch when the physical schema is later validated against the logical schema (e.g., when execution.skip_physical_aggregate_schema_check is false).

To Reproduce

Here is a simple test case to show this behavior.

#[cfg(test)]
mod tests {
    use std::collections::HashMap;
    use std::sync::Arc;

    use datafusion::arrow::datatypes::{DataType, Field, Schema};
    use datafusion::common::ToDFSchema;
    use datafusion::common::metadata::FieldMetadata;
    use datafusion::functions_aggregate::min_max::max_udaf;
    use datafusion::logical_expr::expr::AggregateFunction;
    use datafusion::logical_expr::{EmptyRelation, Expr, LogicalPlan};
    use datafusion::physical_planner::{DefaultPhysicalPlanner, PhysicalPlanner};
    use datafusion::prelude::{SessionContext, col};
    use datafusion_expr::logical_plan::{Aggregate, Projection};

    #[tokio::test]
    async fn test_aggregate_planning_fails_on_projection_with_field_metadata() {
        let schema = Schema::new(vec![Field::new("value", DataType::Utf8, false)]);
        let input = LogicalPlan::EmptyRelation(EmptyRelation {
            produce_one_row: false,
            schema: Arc::new(schema.to_dfschema().unwrap()),
        });
        let metadata = FieldMetadata::from(HashMap::from([("foo".to_string(), "bar".to_string())]));
        let projection = LogicalPlan::Projection(
            Projection::try_new(
                vec![col("value").alias_with_metadata("value", Some(metadata))],
                Arc::new(input),
            )
            .unwrap(),
        );
        let aggregate = LogicalPlan::Aggregate(
            Aggregate::try_new(
                Arc::new(projection),
                vec![],
                vec![Expr::AggregateFunction(AggregateFunction::new_udf(
                    max_udaf(),
                    vec![col("value")],
                    false,
                    None,
                    vec![],
                    None,
                ))],
            )
            .unwrap(),
        );

        DefaultPhysicalPlanner::default()
            .create_physical_plan(&aggregate, &SessionContext::new().state())
            .await
            .unwrap();
    }
}

The test would fail with the following error:

called `Result::unwrap()` on an `Err` value: Internal("Physical input schema should be the same as the one converted from logical input schema. Differences: \n\t- field metadata at index 0 [value]: (physical) {} vs (logical) {\"foo\": \"bar\"}")

Expected behavior

DefaultPhysicalPlanner should handle field metadata in the logical plan, possibly inside create_project_physical_exec_with_props.

Additional context

No response