PIVOT command (#22)

* Pivot initial changes

* Polishing

* Rebase fixes

* Add slt tests

* Add display for PIVOT

* Only check projections in case of subqueries

* Add builder test

* Add default on null. Cargo fmt

* Add default_on_null to proto and tests

* Fix protoc and clippy

* Cargo fmt

Author: Vedin

datafusion/core/src/physical_planner.rs

@@ -76,9 +76,9 @@ use datafusion_expr::expr::{
 use datafusion_expr::expr_rewriter::unnormalize_cols;
 use datafusion_expr::logical_plan::builder::wrap_projection_for_join_if_necessary;
 use datafusion_expr::{
-    Analyze, DescribeTable, DmlStatement, Explain, ExplainFormat, Extension, FetchType,
-    Filter, JoinType, RecursiveQuery, SkipType, StringifiedPlan, WindowFrame,
-    WindowFrameBound, WriteOp,
+    Analyze, BinaryExpr, DescribeTable, DmlStatement, Explain, ExplainFormat, Extension,
+    FetchType, Filter, JoinType, LogicalPlanBuilder, RecursiveQuery, SkipType,
+    StringifiedPlan, WindowFrame, WindowFrameBound, WriteOp,
 };
 use datafusion_physical_expr::aggregate::{AggregateExprBuilder, AggregateFunctionExpr};
 use datafusion_physical_expr::expressions::{Column, Literal};
@@ -91,12 +91,15 @@ use datafusion_physical_plan::unnest::ListUnnest;
 use crate::schema_equivalence::schema_satisfied_by;
 use async_trait::async_trait;
 use datafusion_datasource::file_groups::FileGroup;
+use datafusion_expr_common::operator::Operator;
 use futures::{StreamExt, TryStreamExt};
 use itertools::{multiunzip, Itertools};
 use log::{debug, trace};
 use sqlparser::ast::NullTreatment;
 use tokio::sync::Mutex;
 
+use datafusion_physical_plan::collect;
+
 /// Physical query planner that converts a `LogicalPlan` to an
 /// `ExecutionPlan` suitable for execution.
 #[async_trait]
@@ -887,7 +890,60 @@ impl DefaultPhysicalPlanner {
                     options.clone(),
                 ))
             }
+            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,
@@ -1683,6 +1739,136 @@ pub use datafusion_physical_expr::{
     create_physical_sort_expr, create_physical_sort_exprs,
 };
 
+/// 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: &datafusion_common::Column,
+    pivot_values: Vec<ScalarValue>,
+    default_on_null_expr: Option<&Expr>,
+) -> Result<LogicalPlan> {
+    let df_schema = input.schema();
+
+    let all_columns: Vec<datafusion_common::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: &datafusion_common::Column| {
+            col.name != pivot_column.name
+                && !aggregate_expr
+                    .column_refs()
+                    .iter()
+                    .any(|agg_col| agg_col.name == col.name)
+        })
+        .map(|col: datafusion_common::Column| Expr::Column(col))
+        .collect();
+
+    let builder = LogicalPlanBuilder::from(Arc::unwrap_or_clone(input));
+
+    // Create the aggregate plan with filtered aggregates
+    let mut aggregate_exprs = Vec::new();
+
+    for value in &pivot_values {
+        let filter_condition = Expr::BinaryExpr(BinaryExpr::new(
+            Box::new(Expr::Column(pivot_column.clone())),
+            Operator::IsNotDistinctFrom,
+            Box::new(Expr::Literal(value.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(
+                    datafusion_common::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(datafusion_common::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
     ///
@@ -2044,6 +2230,39 @@ impl DefaultPhysicalPlanner {
             })
             .collect::<Result<Vec<_>>>()?;
 
+        // 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(Column::new(field.name(), i))
+                                    as Arc<dyn PhysicalExpr>,
+                                field.name().clone(),
+                            ));
+                        }
+                    }
+
+                    return Ok(Arc::new(ProjectionExec::try_new(all_exprs, input_exec)?));
+                }
+            }
+        }
+
         Ok(Arc::new(ProjectionExec::try_new(
             physical_exprs,
             input_exec,

datafusion/expr/src/logical_plan/builder.rs

@@ -32,7 +32,7 @@ use crate::expr_rewriter::{
 };
 use crate::logical_plan::{
     Aggregate, Analyze, Distinct, DistinctOn, EmptyRelation, Explain, Filter, Join,
-    JoinConstraint, JoinType, Limit, LogicalPlan, Partitioning, PlanType, Prepare,
+    JoinConstraint, JoinType, Limit, LogicalPlan, Partitioning, Pivot, PlanType, Prepare,
     Projection, Repartition, Sort, SubqueryAlias, TableScan, Union, Unnest, Values,
     Window,
 };
@@ -1427,6 +1427,23 @@ impl LogicalPlanBuilder {
         unnest_with_options(Arc::unwrap_or_clone(self.plan), columns, options)
             .map(Self::new)
     }
+
+    pub fn pivot(
+        self,
+        aggregate_expr: Expr,
+        pivot_column: Column,
+        pivot_values: Vec<ScalarValue>,
+        default_on_null: Option<Expr>,
+    ) -> Result<Self> {
+        let pivot_plan = Pivot::try_new(
+            self.plan,
+            aggregate_expr,
+            pivot_column,
+            pivot_values,
+            default_on_null,
+        )?;
+        Ok(Self::new(LogicalPlan::Pivot(pivot_plan)))
+    }
 }
 
 impl From<LogicalPlan> for LogicalPlanBuilder {
@@ -2824,4 +2841,30 @@ mod tests {
 
         Ok(())
     }
+
+    #[test]
+    fn plan_builder_pivot() -> Result<()> {
+        let schema = Schema::new(vec![
+            Field::new("region", DataType::Utf8, false),
+            Field::new("product", DataType::Utf8, false),
+            Field::new("sales", DataType::Int32, false),
+        ]);
+
+        let plan = LogicalPlanBuilder::scan("sales", table_source(&schema), None)?
+            .pivot(
+                col("sales"),
+                Column::from_name("product"),
+                vec![
+                    ScalarValue::Utf8(Some("widget".to_string())),
+                    ScalarValue::Utf8(Some("gadget".to_string())),
+                ],
+                None,
+            )?
+            .build()?;
+
+        let expected = "Pivot: sales FOR product IN (widget, gadget)\n  TableScan: sales";
+        assert_eq!(expected, format!("{plan}"));
+
+        Ok(())
+    }
 }