feat: add cost estimation for agg (#144)

Compute cost for aggregation. The cardinality is computed as the product
of n-distinct of all the group-by columns. If there's no group by
column, the output cardinality is just 1.

This should fix the cardinality parity between postgres for Q14 and Q17.
It also leads to a better join order in Q11, since aggregation is the
child of a join.

## Misc

- Add planner test for Q11.
- Fixes Q14 and Q17.
- Next step is to support n-distinct for string.
- We may change to multi-dimension n-distinct after it's integrated.

Author: Gun9niR

optd-datafusion-repr/src/cost/base_cost.rs

@@ -1,7 +1,7 @@
 use std::{collections::HashMap, sync::Arc};
 
 use crate::plan_nodes::{
-    BinOpType, ColumnRefExpr, ConstantExpr, ConstantType, LogOpType, OptRelNode, UnOpType,
+    BinOpType, ColumnRefExpr, ConstantExpr, ConstantType, ExprList, LogOpType, OptRelNode, UnOpType,
 };
 use crate::properties::column_ref::{ColumnRefPropertyBuilder, GroupColumnRefs};
 use crate::{
@@ -11,8 +11,8 @@ use crate::{
 use arrow_schema::{ArrowError, DataType};
 use datafusion::arrow::array::{
     Array, BooleanArray, Date32Array, Decimal128Array, Float32Array, Float64Array, Int16Array,
-    Int32Array, Int8Array, RecordBatch, RecordBatchIterator, RecordBatchReader, UInt16Array,
-    UInt32Array, UInt8Array,
+    Int32Array, Int8Array, RecordBatch, RecordBatchIterator, RecordBatchReader, StringArray,
+    UInt16Array, UInt32Array, UInt8Array,
 };
 use itertools::Itertools;
 use optd_core::{
@@ -22,6 +22,7 @@ use optd_core::{
 };
 use optd_gungnir::stats::hyperloglog::{self, HyperLogLog};
 use optd_gungnir::stats::tdigest::{self, TDigest};
+use optd_gungnir::utils::arith_encoder;
 use serde::{Deserialize, Serialize};
 
 fn compute_plan_node_cost<T: RelNodeTyp, C: CostModel<T>>(
@@ -181,6 +182,7 @@ impl DataFusionPerTableStats {
                 | DataType::UInt32
                 | DataType::Float32
                 | DataType::Float64
+                | DataType::Utf8
         )
     }
 
@@ -222,6 +224,10 @@ impl DataFusionPerTableStats {
             val as f64
         }
 
+        fn str_to_f64(string: &str) -> f64 {
+            arith_encoder::encode(string)
+        }
+
         match col_type {
             DataType::Boolean => {
                 generate_stats_for_col!({ col, distr, hll, BooleanArray, to_f64_safe })
@@ -256,6 +262,9 @@ impl DataFusionPerTableStats {
             DataType::Decimal128(_, _) => {
                 generate_stats_for_col!({ col, distr, hll, Decimal128Array, i128_to_f64 })
             }
+            DataType::Utf8 => {
+                generate_stats_for_col!({ col, distr, hll, StringArray, str_to_f64 })
+            }
             _ => unreachable!(),
         }
     }
@@ -323,6 +332,10 @@ const DEFAULT_EQ_SEL: f64 = 0.005;
 const DEFAULT_INEQ_SEL: f64 = 0.3333333333333333;
 // Default selectivity estimate for pattern-match operators such as LIKE
 const DEFAULT_MATCH_SEL: f64 = 0.005;
+// Default selectivity if we have no information
+const DEFAULT_UNK_SEL: f64 = 0.005;
+// Default n-distinct estimate for derived columns or columns lacking statistics
+const DEFAULT_N_DISTINCT: u64 = 200;
 
 const INVALID_SEL: f64 = 0.01;
 
@@ -401,37 +414,33 @@ impl<M: MostCommonValues, D: Distribution> CostModel<OptRelNodeTyp> for OptCostM
             OptRelNodeTyp::PhysicalEmptyRelation => Self::cost(0.5, 0.01, 0.0),
             OptRelNodeTyp::PhysicalLimit => {
                 let (row_cnt, compute_cost, _) = Self::cost_tuple(&children[0]);
-                let row_cnt = if let Some(context) = context {
-                    if let Some(optimizer) = optimizer {
-                        let mut fetch_expr =
-                            optimizer.get_all_group_bindings(context.children_group_ids[2], false);
-                        assert!(
-                            fetch_expr.len() == 1,
-                            "fetch expression should be the only expr in the group"
-                        );
-                        let fetch_expr = fetch_expr.pop().unwrap();
-                        assert!(
-                            matches!(
-                                fetch_expr.typ,
-                                OptRelNodeTyp::Constant(ConstantType::UInt64)
-                            ),
-                            "fetch type can only be UInt64"
-                        );
-                        let fetch = ConstantExpr::from_rel_node(fetch_expr)
-                            .unwrap()
-                            .value()
-                            .as_u64();
-                        // u64::MAX represents None
-                        if fetch == u64::MAX {
-                            row_cnt
-                        } else {
-                            row_cnt.min(fetch as f64)
-                        }
+                let row_cnt = if let (Some(context), Some(optimizer)) = (context, optimizer) {
+                    let mut fetch_expr =
+                        optimizer.get_all_group_bindings(context.children_group_ids[2], false);
+                    assert!(
+                        fetch_expr.len() == 1,
+                        "fetch expression should be the only expr in the group"
+                    );
+                    let fetch_expr = fetch_expr.pop().unwrap();
+                    assert!(
+                        matches!(
+                            fetch_expr.typ,
+                            OptRelNodeTyp::Constant(ConstantType::UInt64)
+                        ),
+                        "fetch type can only be UInt64"
+                    );
+                    let fetch = ConstantExpr::from_rel_node(fetch_expr)
+                        .unwrap()
+                        .value()
+                        .as_u64();
+                    // u64::MAX represents None
+                    if fetch == u64::MAX {
+                        row_cnt
                     } else {
-                        panic!("compute_cost() should not be called if optimizer is None")
+                        row_cnt.min(fetch as f64)
                     }
                 } else {
-                    panic!("compute_cost() should not be called if context is None")
+                    (row_cnt * DEFAULT_UNK_SEL).max(1.0)
                 };
                 Self::cost(row_cnt, compute_cost, 0.0)
             }
@@ -499,10 +508,15 @@ impl<M: MostCommonValues, D: Distribution> CostModel<OptRelNodeTyp> for OptCostM
                 Self::cost(row_cnt, row_cnt * row_cnt.ln_1p().max(1.0), 0.0)
             }
             OptRelNodeTyp::PhysicalAgg => {
-                let (row_cnt, _, _) = Self::cost_tuple(&children[0]);
+                let child_row_cnt = Self::row_cnt(&children[0]);
+                let row_cnt = self.get_agg_row_cnt(context, optimizer, child_row_cnt);
                 let (_, compute_cost_1, _) = Self::cost_tuple(&children[1]);
                 let (_, compute_cost_2, _) = Self::cost_tuple(&children[2]);
-                Self::cost(row_cnt, row_cnt * (compute_cost_1 + compute_cost_2), 0.0)
+                Self::cost(
+                    row_cnt,
+                    child_row_cnt * (compute_cost_1 + compute_cost_2),
+                    0.0,
+                )
             }
             OptRelNodeTyp::List => {
                 let compute_cost = children
@@ -544,6 +558,58 @@ impl<M: MostCommonValues, D: Distribution> OptCostModel<M, D> {
         }
     }
 
+    fn get_agg_row_cnt(
+        &self,
+        context: Option<RelNodeContext>,
+        optimizer: Option<&CascadesOptimizer<OptRelNodeTyp>>,
+        child_row_cnt: f64,
+    ) -> f64 {
+        if let (Some(context), Some(optimizer)) = (context, optimizer) {
+            let group_by_id = context.children_group_ids[2];
+            let mut group_by_exprs: Vec<Arc<RelNode<OptRelNodeTyp>>> =
+                optimizer.get_all_group_bindings(group_by_id, false);
+            assert!(
+                group_by_exprs.len() == 1,
+                "ExprList expression should be the only expression in the GROUP BY group"
+            );
+            let group_by = group_by_exprs.pop().unwrap();
+            let group_by = ExprList::from_rel_node(group_by).unwrap();
+            if group_by.is_empty() {
+                1.0
+            } else {
+                // Multiply the n-distinct of all the group by columns.
+                // TODO: improve with multi-dimensional n-distinct
+                let base_table_col_refs = optimizer
+                    .get_property_by_group::<ColumnRefPropertyBuilder>(context.group_id, 1);
+                base_table_col_refs
+                    .iter()
+                    .take(group_by.len())
+                    .map(|col_ref| match col_ref {
+                        ColumnRef::BaseTableColumnRef { table, col_idx } => {
+                            let table_stats = self.per_table_stats_map.get(table);
+                            let column_stats = table_stats.map(|table_stats| {
+                                table_stats.per_column_stats_vec.get(*col_idx).unwrap()
+                            });
+
+                            if let Some(Some(column_stats)) = column_stats {
+                                column_stats.ndistinct as f64
+                            } else {
+                                // The column type is not supported or stats are missing.
+                                DEFAULT_N_DISTINCT as f64
+                            }
+                        }
+                        ColumnRef::Derived => DEFAULT_N_DISTINCT as f64,
+                        _ => panic!(
+                            "GROUP BY base table column ref must either be derived or base table"
+                        ),
+                    })
+                    .product()
+            }
+        } else {
+            (child_row_cnt * DEFAULT_UNK_SEL).max(1.0)
+        }
+    }
+
     /// The expr_tree input must be a "mixed expression tree"
     ///     An "expression node" refers to a RelNode that returns true for is_expression()
     ///     A "full expression tree" is where every node in the tree is an expression node

optd-datafusion-repr/src/properties/schema.rs

@@ -4,14 +4,26 @@ use std::sync::Arc;
 use optd_core::property::PropertyBuilder;
 
 use super::DEFAULT_NAME;
-use crate::plan_nodes::{ConstantType, EmptyRelationData, OptRelNodeTyp};
+use crate::plan_nodes::{ConstantType, EmptyRelationData, FuncType, OptRelNodeTyp};
 
 #[derive(Clone, Debug, Serialize, Deserialize)]
 pub struct Field {
     pub name: String,
     pub typ: ConstantType,
     pub nullable: bool,
 }
+
+impl Field {
+    /// Generate a field that is only a place holder whose members are never used.
+    fn placeholder() -> Self {
+        Self {
+            name: DEFAULT_NAME.to_string(),
+            typ: ConstantType::Any,
+            nullable: true,
+        }
+    }
+}
+
 #[derive(Clone, Debug, Serialize, Deserialize)]
 pub struct Schema {
     pub fields: Vec<Field>,
@@ -87,14 +99,18 @@ impl PropertyBuilder<OptRelNodeTyp> for SchemaPropertyBuilder {
                 Schema { fields }
             }
             OptRelNodeTyp::LogOp(_) => Schema {
-                fields: vec![
-                    Field {
-                        name: DEFAULT_NAME.to_string(),
-                        typ: ConstantType::Any,
-                        nullable: true
-                    };
-                    children.len()
-                ],
+                fields: vec![Field::placeholder(); children.len()],
+            },
+            OptRelNodeTyp::Agg => {
+                let mut group_by_schema = children[1].clone();
+                let agg_schema = children[2].clone();
+                group_by_schema.fields.extend(agg_schema.fields);
+                group_by_schema
+            }
+            OptRelNodeTyp::Func(FuncType::Agg(_)) => Schema {
+                // TODO: this is just a place holder now.
+                // The real type should be the column type.
+                fields: vec![Field::placeholder()],
             },
             _ => Schema { fields: vec![] },
         }

optd-gungnir/src/stats/hyperloglog.rs

@@ -25,10 +25,17 @@ pub struct HyperLogLog {
     alpha: f64,         // The normal HLL multiplier factor.
 }
 
+// Serialize common data types for hashing (&str).
+impl ByteSerializable for &str {
+    fn to_bytes(&self) -> Vec<u8> {
+        self.as_bytes().to_vec()
+    }
+}
+
 // Serialize common data types for hashing (String).
 impl ByteSerializable for String {
     fn to_bytes(&self) -> Vec<u8> {
-        self.as_bytes().to_vec()
+        self.as_str().to_bytes()
     }
 }
 

optd-sqlplannertest/src/lib.rs

@@ -140,7 +140,6 @@ impl DatafusionDBMS {
         task: &str,
         flags: &[String],
     ) -> Result<()> {
-        println!("task_explain(): called on sql={}", sql);
         use std::fmt::Write;
 
         let with_logical = flags.contains(&"with_logical".to_string());