refactor(core): drop value-bag dependency for statistics (#227)

value-bag was added for easy ser/deserialization of the statistics. However, after second thought, this wouldn't be possible to be compatible with serde API: when someone wants to deserialize something, they need to know the underlying type to deserialize, and this info is erased behind `dyn`. Therefore, the correct approach to enable serialization for the properties is to have two new methods on the PropertyBuilder / CostModel trait: `serialize_stats(&self, &dyn Any) -> serde_json::Value` and `deserialize_stats(&self, serde_json::Value) -> Box<dyn Any>`. --------- Signed-off-by: Alex Chi <[email protected]>
cmu-db · Nov 8, 2024 · 75f9fe6 · 75f9fe6
1 parent 4f84645
commit 75f9fe6
Show file tree

Hide file tree

Showing 8 changed files with 23 additions and 156 deletions.
diff --git a/Cargo.lock b/Cargo.lock
diff --git a/optd-core/Cargo.toml b/optd-core/Cargo.toml
@@ -18,4 +18,4 @@ itertools = "0.13"
 serde = { version = "1.0", features = ["derive", "rc"] }
 arrow-schema = "47.0.0"
 chrono = "0.4"
-value-bag = { version = "1", features = ["owned"] }
+erased-serde = "0.4"
diff --git a/optd-core/src/cascades/memo.rs b/optd-core/src/cascades/memo.rs
@@ -41,7 +41,7 @@ impl<T: NodeType> std::fmt::Display for MemoPlanNode<T> {
     }
 }
 
-#[derive(Debug, Clone)]
+#[derive(Clone)]
 pub struct WinnerInfo {
     pub expr_id: ExprId,
     pub total_weighted_cost: f64,
@@ -51,7 +51,7 @@ pub struct WinnerInfo {
     pub statistics: Arc<Statistics>,
 }
 
-#[derive(Debug, Clone)]
+#[derive(Clone)]
 pub enum Winner {
     Unknown,
     Impossible,
@@ -81,7 +81,7 @@ impl Default for Winner {
     }
 }
 
-#[derive(Default, Debug, Clone)]
+#[derive(Default, Clone)]
 pub struct GroupInfo {
     pub winner: Winner,
 }

diff --git a/optd-core/src/cost.rs b/optd-core/src/cost.rs
@@ -7,8 +7,7 @@ use crate::cascades::{CascadesOptimizer, Memo, RelNodeContext};
 use crate::nodes::{ArcPredNode, NodeType};
 
 /// The statistics of a group.
-#[derive(Clone, Debug)]
-pub struct Statistics(pub value_bag::OwnedValueBag);
+pub struct Statistics(pub Box<dyn std::any::Any + Send + Sync + 'static>);
 
 /// The cost of an operation. The cost is represented as a vector of double values.
 /// For example, it can be represented as `[compute_cost, io_cost]`.

diff --git a/optd-core/src/nodes.rs b/optd-core/src/nodes.rs
@@ -358,7 +358,7 @@ impl<T: NodeType> PredNode<T> {
 }
 
 /// Metadata for a rel node.
-#[derive(Clone, Debug)]
+#[derive(Clone)]
 pub struct PlanNodeMeta {
     /// The group (id) of the `RelNode`
     pub group_id: GroupId,

diff --git a/optd-datafusion-repr/Cargo.toml b/optd-datafusion-repr/Cargo.toml
@@ -22,4 +22,3 @@ camelpaste = "0.1"
 datafusion-expr = "32.0.0"
 serde = { version = "1.0", features = ["derive"] }
 bincode = "1.3.3"
-value-bag = { version = "1", features = ["owned"] }
diff --git a/optd-datafusion-repr/src/cost/base_cost.rs b/optd-datafusion-repr/src/cost/base_cost.rs
@@ -8,10 +8,14 @@ use std::collections::HashMap;
 use itertools::Itertools;
 use optd_core::cascades::{CascadesOptimizer, NaiveMemo, RelNodeContext};
 use optd_core::cost::{Cost, CostModel, Statistics};
-use value_bag::ValueBag;
 
 use crate::plan_nodes::{ArcDfPredNode, ConstantPred, DfNodeType, DfReprPredNode};
 
+#[derive(Debug, Clone)]
+pub struct DfStatistics {
+    row_cnt: f64,
+}
+
 pub struct DfCostModel {
     table_stat: HashMap<String, usize>,
 }
@@ -31,15 +35,15 @@ impl DfCostModel {
     }
 
     pub fn row_cnt(Statistics(stat): &Statistics) -> f64 {
-        stat.by_ref().as_f64()
+        stat.downcast_ref::<DfStatistics>().unwrap().row_cnt
     }
 
     pub fn cost(compute_cost: f64, io_cost: f64) -> Cost {
         Cost(vec![compute_cost, io_cost])
     }
 
     pub fn stat(row_cnt: f64) -> Statistics {
-        Statistics(ValueBag::from_f64(row_cnt).to_owned())
+        Statistics(Box::new(DfStatistics { row_cnt }))
     }
 
     pub fn cost_tuple(Cost(cost): &Cost) -> (f64, f64) {

diff --git a/optd-datafusion-repr/src/testing/dummy_cost.rs b/optd-datafusion-repr/src/testing/dummy_cost.rs
@@ -5,7 +5,6 @@
 
 use optd_core::cascades::{CascadesOptimizer, NaiveMemo, RelNodeContext};
 use optd_core::cost::{Cost, CostModel, Statistics};
-use value_bag::ValueBag;
 
 use crate::plan_nodes::{ArcDfPredNode, DfNodeType};
 
@@ -36,7 +35,7 @@ impl CostModel<DfNodeType, NaiveMemo<DfNodeType>> for DummyCostModel {
         _: Option<RelNodeContext>,
         _: Option<&CascadesOptimizer<DfNodeType>>,
     ) -> Statistics {
-        Statistics(ValueBag::empty().to_owned())
+        Statistics(Box::new(()))
     }
 
     fn explain_cost(&self, _: &Cost) -> String {