Fixes after sync to 50.0.0

Author: Vedin

datafusion/core/src/execution/session_state.rs

@@ -292,7 +292,9 @@ impl Session for SessionState {
 }
 
 impl SessionState {
-    pub(crate) fn resolve_table_ref(
+    /// Resolve a [`TableReference`] into a [`ResolvedTableReference`] using
+    /// the session's configured default catalog and schema.
+    pub fn resolve_table_ref(
         &self,
         table_ref: impl Into<TableReference>,
     ) -> ResolvedTableReference {

datafusion/core/src/physical_planner.rs

@@ -86,12 +86,12 @@ use datafusion_expr::expr_rewriter::unnormalize_cols;
 use datafusion_expr::logical_plan::builder::wrap_projection_for_join_if_necessary;
 use datafusion_expr::utils::{expr_to_columns, split_conjunction};
 use datafusion_expr::{
-    Analyze, BinaryExpr, DescribeTable, DmlStatement, Explain, ExplainFormat, Extension,
-    FetchType, Filter, JoinType, Operator, RecursiveQuery, SkipType, StringifiedPlan,
-    WindowFrame, WindowFrameBound, WriteOp,
+    Analyze, BinaryExpr, Cast, DescribeTable, DmlStatement, Explain, ExplainFormat,
+    Extension, FetchType, Filter, JoinType, LogicalPlanBuilder, Operator, RecursiveQuery,
+    SkipType, StringifiedPlan, WindowFrame, WindowFrameBound, WriteOp,
 };
 use datafusion_physical_expr::aggregate::{AggregateExprBuilder, AggregateFunctionExpr};
-use datafusion_physical_expr::expressions::Literal;
+use datafusion_physical_expr::expressions::{Column as PhysicalColumn, Literal};
 use datafusion_physical_expr::{
     LexOrdering, PhysicalSortExpr, create_physical_sort_exprs,
 };
@@ -106,6 +106,7 @@ use datafusion_physical_plan::unnest::ListUnnest;
 
 use async_trait::async_trait;
 use datafusion_physical_plan::async_func::{AsyncFuncExec, AsyncMapper};
+use datafusion_physical_plan::collect;
 use futures::{StreamExt, TryStreamExt};
 use itertools::{Itertools, multiunzip};
 use log::debug;
@@ -1256,6 +1257,56 @@ impl DefaultPhysicalPlanner {
                 )?)
             }
 
+            LogicalPlan::Pivot(pivot) => {
+                return if !pivot.pivot_values.is_empty() {
+                    let agg_plan = transform_pivot_to_aggregate(
+                        Arc::new(pivot.input.as_ref().clone()),
+                        &pivot.aggregate_expr,
+                        &pivot.pivot_column,
+                        pivot.pivot_values.clone(),
+                        pivot.default_on_null_expr.as_ref(),
+                    )?;
+                    self.create_physical_plan(&agg_plan, session_state).await
+                } else if let Some(subquery) = &pivot.value_subquery {
+                    let optimized_subquery = session_state.optimize(subquery.as_ref())?;
+                    let subquery_physical_plan = self
+                        .create_physical_plan(&optimized_subquery, session_state)
+                        .await?;
+                    let subquery_results = collect(
+                        Arc::clone(&subquery_physical_plan),
+                        session_state.task_ctx(),
+                    )
+                    .await?;
+
+                    let mut pivot_values = Vec::new();
+                    for batch in subquery_results.iter() {
+                        if batch.num_columns() != 1 {
+                            return plan_err!(
+                                "Pivot subquery must return a single column"
+                            );
+                        }
+                        let column = batch.column(0);
+                        for row_idx in 0..batch.num_rows() {
+                            if !column.is_null(row_idx) {
+                                pivot_values
+                                    .push(ScalarValue::try_from_array(column, row_idx)?);
+                            }
+                        }
+                    }
+
+                    let agg_plan = transform_pivot_to_aggregate(
+                        Arc::new(pivot.input.as_ref().clone()),
+                        &pivot.aggregate_expr,
+                        &pivot.pivot_column,
+                        pivot_values,
+                        pivot.default_on_null_expr.as_ref(),
+                    )?;
+                    self.create_physical_plan(&agg_plan, session_state).await
+                } else {
+                    plan_err!("PIVOT operation requires at least one value to pivot on")
+                };
+            }
+
             // 2 Children
             LogicalPlan::Join(Join {
                 left: original_left,
@@ -2143,7 +2194,8 @@ fn extract_dml_filters(
             | LogicalPlan::Ddl(_)
             | LogicalPlan::Copy(_)
             | LogicalPlan::Unnest(_)
-            | LogicalPlan::RecursiveQuery(_) => {
+            | LogicalPlan::RecursiveQuery(_)
+            | LogicalPlan::Pivot(_) => {
                 // No filters to extract from leaf/meta plans
             }
             // Plans with inputs (may contain filters in children)
@@ -2499,6 +2551,147 @@ pub fn create_aggregate_expr_and_maybe_filter(
     )
 }
 
+/// Transform a PIVOT operation into a more standard Aggregate + Projection plan
+/// For known pivot values, we create a projection that includes "IS NOT DISTINCT FROM" conditions
+///
+/// For example, for SUM(amount) PIVOT(quarter FOR quarter in ('2023_Q1', '2023_Q2')), we create:
+/// - SUM(amount) FILTER (WHERE quarter IS NOT DISTINCT FROM '2023_Q1') AS "2023_Q1"
+/// - SUM(amount) FILTER (WHERE quarter IS NOT DISTINCT FROM '2023_Q2') AS "2023_Q2"
+///
+/// If DEFAULT ON NULL is specified, each aggregate expression is wrapped with an outer projection that
+/// applies COALESCE to the results.
+pub fn transform_pivot_to_aggregate(
+    input: Arc<LogicalPlan>,
+    aggregate_expr: &Expr,
+    pivot_column: &Column,
+    pivot_values: Vec<ScalarValue>,
+    default_on_null_expr: Option<&Expr>,
+) -> Result<LogicalPlan> {
+    let df_schema = input.schema();
+
+    let all_columns: Vec<Column> = df_schema.columns();
+
+    // Filter to include only columns we want for GROUP BY
+    // (exclude pivot column and aggregate expression columns)
+    let group_by_columns: Vec<Expr> = all_columns
+        .into_iter()
+        .filter(|col: &Column| {
+            col.name != pivot_column.name
+                && !aggregate_expr
+                    .column_refs()
+                    .iter()
+                    .any(|agg_col| agg_col.name == col.name)
+        })
+        .map(|col: Column| Expr::Column(col))
+        .collect();
+
+    let builder = LogicalPlanBuilder::from(Arc::unwrap_or_clone(Arc::clone(&input)));
+
+    // Create the aggregate plan with filtered aggregates
+    let mut aggregate_exprs = Vec::new();
+
+    let input_schema = input.schema();
+    let pivot_col_idx = match input_schema.index_of_column(pivot_column) {
+        Ok(idx) => idx,
+        Err(_) => {
+            return plan_err!(
+                "Pivot column '{}' does not exist in input schema",
+                pivot_column
+            );
+        }
+    };
+    let pivot_col_type = input_schema.field(pivot_col_idx).data_type();
+
+    for value in &pivot_values {
+        let filter_condition = Expr::BinaryExpr(BinaryExpr::new(
+            Box::new(Expr::Column(pivot_column.clone())),
+            Operator::IsNotDistinctFrom,
+            Box::new(Expr::Cast(Cast::new(
+                Box::new(Expr::Literal(value.clone(), None)),
+                pivot_col_type.clone(),
+            ))),
+        ));
+
+        let filtered_agg = match aggregate_expr {
+            Expr::AggregateFunction(agg) => {
+                let mut new_params = agg.params.clone();
+                new_params.filter = Some(Box::new(filter_condition));
+                Expr::AggregateFunction(AggregateFunction {
+                    func: Arc::clone(&agg.func),
+                    params: new_params,
+                })
+            }
+            _ => {
+                return plan_err!(
+                    "Unsupported aggregate expression should always be AggregateFunction"
+                );
+            }
+        };
+
+        // Use the pivot value as the column name
+        let field_name = value.to_string().trim_matches('\'').to_string();
+        let aliased_agg = Expr::Alias(Alias {
+            expr: Box::new(filtered_agg),
+            relation: None,
+            name: field_name,
+            metadata: None,
+        });
+
+        aggregate_exprs.push(aliased_agg);
+    }
+
+    // Create the plan with the aggregate
+    let aggregate_plan = builder
+        .aggregate(group_by_columns, aggregate_exprs)?
+        .build()?;
+
+    // If DEFAULT ON NULL is specified, add a projection to apply COALESCE
+    if let Some(default_expr) = default_on_null_expr {
+        let schema = aggregate_plan.schema();
+        let mut projection_exprs = Vec::new();
+
+        for field in schema.fields() {
+            if !pivot_values
+                .iter()
+                .any(|v| field.name() == v.to_string().trim_matches('\''))
+            {
+                projection_exprs.push(Expr::Column(Column::from_name(field.name())));
+            }
+        }
+
+        // Apply COALESCE to aggregate columns
+        for value in &pivot_values {
+            let field_name = value.to_string().trim_matches('\'').to_string();
+            let aggregate_col = Expr::Column(Column::from_name(&field_name));
+
+            // Create COALESCE expression using CASE: CASE WHEN col IS NULL THEN default_value ELSE col END
+            let coalesce_expr = Expr::Case(datafusion_expr::expr::Case {
+                expr: None,
+                when_then_expr: vec![(
+                    Box::new(Expr::IsNull(Box::new(aggregate_col.clone()))),
+                    Box::new(default_expr.clone()),
+                )],
+                else_expr: Some(Box::new(aggregate_col)),
+            });
+
+            let aliased_coalesce = Expr::Alias(Alias {
+                expr: Box::new(coalesce_expr),
+                relation: None,
+                name: field_name,
+                metadata: None,
+            });
+
+            projection_exprs.push(aliased_coalesce);
+        }
+
+        // Apply the projection
+        LogicalPlanBuilder::from(aggregate_plan)
+            .project(projection_exprs)?
+            .build()
+    } else {
+        Ok(aggregate_plan)
+    }
+}
 impl DefaultPhysicalPlanner {
     /// Handles capturing the various plans for EXPLAIN queries
     ///
@@ -2881,6 +3074,38 @@ impl DefaultPhysicalPlanner {
             .collect::<Result<Vec<_>>>()?;
 
         let num_input_columns = input_exec.schema().fields().len();
+        // When we detect a PIVOT-derived plan with a value_subquery, ensure all generated columns are preserved
+        if let LogicalPlan::Pivot(pivot) = input.as_ref() {
+            if pivot.value_subquery.is_some()
+                && input_exec
+                    .as_any()
+                    .downcast_ref::<AggregateExec>()
+                    .is_some()
+            {
+                let agg_exec =
+                    input_exec.as_any().downcast_ref::<AggregateExec>().unwrap();
+                let schema = input_exec.schema();
+                let group_by_len = agg_exec.group_expr().expr().len();
+
+                if group_by_len < schema.fields().len() {
+                    let mut all_exprs = physical_exprs.clone();
+
+                    for (i, field) in
+                        schema.fields().iter().enumerate().skip(group_by_len)
+                    {
+                        if !physical_exprs.iter().any(|(_, name)| name == field.name()) {
+                            all_exprs.push((
+                                Arc::new(PhysicalColumn::new(field.name(), i))
+                                    as Arc<dyn PhysicalExpr>,
+                                field.name().clone(),
+                            ));
+                        }
+                    }
+
+                    return Ok(Arc::new(ProjectionExec::try_new(all_exprs, input_exec)?));
+                }
+            }
+        }
 
         match self.try_plan_async_exprs(
             num_input_columns,

datafusion/expr/src/logical_plan/display.rs

@@ -313,6 +313,12 @@ impl<'a, 'b> PgJsonVisitor<'a, 'b> {
                     "Is Distinct": is_distinct,
                 })
             }
+            LogicalPlan::Pivot(pivot) => json!({
+                "Node Type": "Pivot",
+                "Aggregate Expr": format!("{}", pivot.aggregate_expr),
+                "Pivot Column": format!("{}", pivot.pivot_column),
+                "Pivot Values": pivot.pivot_values.iter().map(|v| v.to_string()).collect::<Vec<_>>(),
+            }),
             LogicalPlan::Values(Values { values, .. }) => {
                 let str_values = values
                     .iter()
@@ -706,8 +712,11 @@ impl<'n> TreeNodeVisitor<'n> for PgJsonVisitor<'_, '_> {
 
 #[cfg(test)]
 mod tests {
-    use arrow::datatypes::{DataType, Field};
+    use crate::EmptyRelation;
+    use arrow::datatypes::{DataType, Field, Schema};
+    use datafusion_common::{Column, DFSchema, ScalarValue};
     use insta::assert_snapshot;
+    use std::sync::Arc;
 
     use super::*;
 

datafusion/expr/src/logical_plan/mod.rs

@@ -40,8 +40,8 @@ pub use dml::{DmlStatement, WriteOp};
 pub use plan::{
     Aggregate, Analyze, ColumnUnnestList, DescribeTable, Distinct, DistinctOn,
     EmptyRelation, Explain, ExplainOption, Extension, FetchType, Filter, Join,
-    JoinConstraint, JoinType, Limit, LogicalPlan, Partitioning, PlanType, Projection,
-    RecursiveQuery, Repartition, SkipType, Sort, StringifiedPlan, Subquery,
+    JoinConstraint, JoinType, Limit, LogicalPlan, Partitioning, Pivot, PlanType,
+    Projection, RecursiveQuery, Repartition, SkipType, Sort, StringifiedPlan, Subquery,
     SubqueryAlias, TableScan, ToStringifiedPlan, Union, Unnest, Values, Window,
     projection_schema,
 };