Merge branch 'temp' into reclaimer-join-reorder

Author: JanKaul

datafusion/catalog/src/default_table_source.rs

@@ -23,7 +23,7 @@ use std::sync::Arc;
 use crate::TableProvider;
 
 use arrow::datatypes::SchemaRef;
-use datafusion_common::{Constraints, internal_err};
+use datafusion_common::{Constraints, Statistics, internal_err};
 use datafusion_expr::{Expr, TableProviderFilterPushDown, TableSource, TableType};
 
 /// Implements [`TableSource`] for a [`TableProvider`]
@@ -77,6 +77,10 @@ impl TableSource for DefaultTableSource {
     fn get_column_default(&self, column: &str) -> Option<&Expr> {
         self.table_provider.get_column_default(column)
     }
+
+    fn statistics(&self) -> Option<Statistics> {
+        self.table_provider.statistics()
+    }
 }
 
 /// Wrap TableProvider in TableSource

datafusion/expr/src/table_source.rs

@@ -20,7 +20,7 @@
 use crate::{Expr, LogicalPlan};
 
 use arrow::datatypes::SchemaRef;
-use datafusion_common::{Constraints, Result};
+use datafusion_common::{Constraints, Result, Statistics};
 
 use std::{any::Any, borrow::Cow};
 
@@ -127,6 +127,14 @@ pub trait TableSource: Any + Sync + Send {
     fn get_column_default(&self, _column: &str) -> Option<&Expr> {
         None
     }
+
+    /// Get statistics for this table, if available
+    /// Although not presently used in mainline DataFusion, this allows implementation specific
+    /// behavior for downstream repositories, in conjunction with specialized optimizer rules to
+    /// perform operations such as re-ordering of joins.
+    fn statistics(&self) -> Option<Statistics> {
+        None
+    }
 }
 
 impl dyn TableSource {

datafusion/optimizer/src/lib.rs

@@ -65,6 +65,7 @@ pub mod optimizer;
 pub mod propagate_empty_relation;
 pub mod push_down_filter;
 pub mod push_down_limit;
+pub mod reorder_join;
 pub mod replace_distinct_aggregate;
 pub mod rewrite_set_comparison;
 pub mod scalar_subquery_to_join;

datafusion/optimizer/src/reorder_join/cost.rs

@@ -0,0 +1,245 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+use datafusion_common::{Column, Result, plan_err, stats::Precision};
+use datafusion_expr::{Expr, JoinType, LogicalPlan};
+
+use super::join_graph::Edge;
+
+/// Fraction of preserved-side rows estimated to survive a semi/anti join
+/// when column NDV statistics are unavailable. Mirrors DuckDB's
+/// `CardinalityEstimator::DEFAULT_SEMI_ANTI_SELECTIVITY = 1/5`.
+const DEFAULT_SEMI_ANTI_SELECTIVITY: f64 = 0.2;
+
+pub trait JoinCostEstimator: std::fmt::Debug {
+    /// Cardinality of `plan`.
+    ///
+    /// - `column = None`: number of output rows of `plan`.
+    /// - `column = Some(c)`: number of distinct values of column `c`
+    ///   in `plan`'s output (NDV).
+    fn cardinality(&self, plan: &LogicalPlan, column: Option<&Column>) -> Option<f64> {
+        estimate_cardinality(plan, column).ok()
+    }
+
+    /// Estimated selectivity of joining `left` with `right` via `edge`.
+    ///
+    /// Default: `1 / max(NDV(left.key), NDV(right.key))` for equi-joins
+    /// (inner and semi/anti) when both NDVs are available; otherwise a
+    /// per-join-type constant.
+    fn selectivity(&self, edge: &Edge, left: &LogicalPlan, right: &LogicalPlan) -> f64 {
+        let fallback = match edge.join_type {
+            JoinType::Inner => 0.1,
+            JoinType::LeftSemi
+            | JoinType::LeftAnti
+            | JoinType::RightSemi
+            | JoinType::RightAnti => DEFAULT_SEMI_ANTI_SELECTIVITY,
+            _ => 1.0,
+        };
+        let is_eq_join = matches!(
+            edge.join_type,
+            JoinType::Inner
+                | JoinType::LeftSemi
+                | JoinType::LeftAnti
+                | JoinType::RightSemi
+                | JoinType::RightAnti
+        );
+        if !is_eq_join || edge.on.is_empty() {
+            return fallback;
+        }
+        // Use only the first equi-pair. Compounding pairwise selectivities
+        // under independence assumptions overestimates selectivity when
+        // composite-key columns are correlated, which is the common case.
+        let (a, b) = &edge.on[0];
+        let (Some(col_a), Some(col_b)) = (key_column(a), key_column(b)) else {
+            return fallback;
+        };
+        let ndv_a = ndv_for(self, col_a, left, right);
+        let ndv_b = ndv_for(self, col_b, left, right);
+        match edge.join_type {
+            JoinType::Inner => match (ndv_a, ndv_b) {
+                (Some(a), Some(b)) if a.max(b) > 0.0 => 1.0 / a.max(b),
+                _ => fallback,
+            },
+            // Semi/anti containment estimator: surviving fraction of the
+            // preserved side ≈ `min(NDV_preserved, NDV_filtering) / NDV_preserved`.
+            // Edges normalized by `flatten_joins_recursive` always have
+            // `on = (preserved_key, filtering_key)`, so the preserved
+            // NDV is `ndv_a` for Left{Semi,Anti}. RightSemi/RightAnti
+            // shouldn't appear in graph edges (they get normalized) but
+            // are handled defensively.
+            JoinType::LeftSemi | JoinType::LeftAnti => match (ndv_a, ndv_b) {
+                (Some(a), Some(b)) if a > 0.0 => (a.min(b) / a).min(1.0),
+                _ => fallback,
+            },
+            JoinType::RightSemi | JoinType::RightAnti => match (ndv_a, ndv_b) {
+                (Some(a), Some(b)) if b > 0.0 => (a.min(b) / b).min(1.0),
+                _ => fallback,
+            },
+            _ => fallback,
+        }
+    }
+
+    fn cost(&self, selectivity: f64, cardinality: f64) -> f64 {
+        selectivity * cardinality
+    }
+}
+
+/// Default implementation of JoinCostEstimator
+#[derive(Debug, Clone, Copy)]
+pub struct DefaultCostEstimator;
+
+impl JoinCostEstimator for DefaultCostEstimator {}
+
+fn key_column(expr: &Expr) -> Option<&Column> {
+    match expr {
+        Expr::Column(c) => Some(c),
+        _ => None,
+    }
+}
+
+/// Look up NDV of `column` on whichever side (left or right) owns it.
+fn ndv_for<E: JoinCostEstimator + ?Sized>(
+    estimator: &E,
+    column: &Column,
+    left: &LogicalPlan,
+    right: &LogicalPlan,
+) -> Option<f64> {
+    if left.schema().has_column(column) {
+        estimator.cardinality(left, Some(column))
+    } else if right.schema().has_column(column) {
+        estimator.cardinality(right, Some(column))
+    } else {
+        None
+    }
+}
+
+fn estimate_cardinality(plan: &LogicalPlan, column: Option<&Column>) -> Result<f64> {
+    match plan {
+        LogicalPlan::Filter(filter) => match column {
+            None => {
+                let input = estimate_cardinality(&filter.input, None)?;
+                Ok(0.1 * input)
+            }
+            Some(c) => {
+                // NDV is bounded above by the input's NDV and by the
+                // surviving row count.
+                let ndv_in = estimate_cardinality(&filter.input, Some(c))?;
+                let rows_out = estimate_cardinality(plan, None).unwrap_or(ndv_in);
+                Ok(ndv_in.min(rows_out))
+            }
+        },
+        LogicalPlan::Aggregate(agg) => match column {
+            None => {
+                let input = estimate_cardinality(&agg.input, None)?;
+                Ok(0.1 * input)
+            }
+            Some(c) => {
+                // Group-by keys are unique in the aggregate's output, so
+                // NDV(group_key) equals the post-aggregate row count.
+                let is_group_key = agg.group_expr.iter().any(|e| match e {
+                    Expr::Column(g) => g.name == c.name && g.relation == c.relation,
+                    _ => false,
+                });
+                if is_group_key {
+                    estimate_cardinality(plan, None)
+                } else {
+                    plan_err!(
+                        "Cannot estimate NDV of non-group-by column \
+                         `{}` through Aggregate",
+                        c.name
+                    )
+                }
+            }
+        },
+        LogicalPlan::TableScan(scan) => {
+            let stats = scan.source.statistics().ok_or_else(|| {
+                datafusion_common::DataFusionError::Plan(format!(
+                    "TableSource for `{}` does not expose statistics",
+                    scan.table_name
+                ))
+            })?;
+            match column {
+                None => match stats.num_rows {
+                    Precision::Exact(n) | Precision::Inexact(n) => Ok(n as f64),
+                    Precision::Absent => plan_err!(
+                        "TableSource for `{}` does not provide a row count",
+                        scan.table_name
+                    ),
+                },
+                Some(c) => {
+                    // `column_statistics` is indexed by the source schema
+                    // (pre-projection), so resolve the column there.
+                    let idx = scan.source.schema().index_of(&c.name).map_err(|_| {
+                        datafusion_common::DataFusionError::Plan(format!(
+                            "Column `{}` not found in source schema of `{}`",
+                            c.name, scan.table_name
+                        ))
+                    })?;
+                    let col_stats =
+                        stats.column_statistics.get(idx).ok_or_else(|| {
+                            datafusion_common::DataFusionError::Plan(format!(
+                                "Column statistics missing for index {idx} \
+                                 on `{}`",
+                                scan.table_name
+                            ))
+                        })?;
+                    match col_stats.distinct_count {
+                        Precision::Exact(n) | Precision::Inexact(n) => Ok(n as f64),
+                        Precision::Absent => plan_err!(
+                            "Column `{}` on `{}` has no distinct-count statistic",
+                            c.name,
+                            scan.table_name
+                        ),
+                    }
+                }
+            }
+        }
+        // Semi/anti joins do not grow rows: the output cardinality is
+        // bounded by the preserved side. We size them via the
+        // `DEFAULT_SEMI_ANTI_SELECTIVITY` heuristic. NDV queries on the
+        // output route to whichever side is preserved.
+        LogicalPlan::Join(j)
+            if matches!(
+                j.join_type,
+                JoinType::LeftSemi
+                    | JoinType::LeftAnti
+                    | JoinType::RightSemi
+                    | JoinType::RightAnti
+            ) =>
+        {
+            let preserved = match j.join_type {
+                JoinType::LeftSemi | JoinType::LeftAnti => &j.left,
+                _ => &j.right,
+            };
+            match column {
+                None => {
+                    let rows = estimate_cardinality(preserved, None)?;
+                    Ok(rows * DEFAULT_SEMI_ANTI_SELECTIVITY)
+                }
+                Some(c) => estimate_cardinality(preserved, Some(c)),
+            }
+        }
+        x => {
+            let inputs = x.inputs();
+            if inputs.len() == 1 {
+                estimate_cardinality(inputs[0], column)
+            } else {
+                plan_err!("Cannot estimate cardinality for plan with multiple inputs")
+            }
+        }
+    }
+}