opt: make placeholder fast path conditional on the estimated row count

This change is best explained by this comment:

```
// We are dealing with a memo that still contains placeholders. The statistics
// for such a memo can be wildly overestimated. Even if our plan is correct,
// the estimated row count for a scan is passed to the execution engine which
// uses it to make sizing decisions. Passing a very high count can affect
// performance significantly (see #64214). So we only use the fast path if the
// estimated row count is small; typically this will happen when we constrain
// columns that form a key and we know there will be at most one row.
```

Fixes #64214.

Release note (bug fix): fixed a performance regression for very simple
queries.

pkg/sql/opt/xform/placeholder_fast_path.go

@@ -20,6 +20,8 @@ import (
 	"github.com/cockroachdb/errors"
 )
 
+const maxRowCountForPlaceholderFastPath = 10
+
 // TryPlaceholderFastPath attempts to produce a fully optimized memo with
 // placeholders. This is only possible in very simple cases and involves special
 // operators (PlaceholderScan) which use placeholders and resolve them at
@@ -47,13 +49,24 @@ func (o *Optimizer) TryPlaceholderFastPath() (_ opt.Expr, ok bool, err error) {
 	}()
 
 	root := o.mem.RootExpr().(memo.RelExpr)
-	rootProps := o.mem.RootProps()
 
-	if !rootProps.Ordering.Any() {
+	rootRelProps := root.Relational()
+	// We are dealing with a memo that still contains placeholders. The statistics
+	// for such a memo can be wildly overestimated. Even if our plan is correct,
+	// the estimated row count for a scan is passed to the execution engine which
+	// uses it to make sizing decisions. Passing a very high count can affect
+	// performance significantly (see #64214). So we only use the fast path if the
+	// estimated row count is small; typically this will happen when we constrain
+	// columns that form a key and we know there will be at most one row.
+	if rootRelProps.Stats.RowCount > maxRowCountForPlaceholderFastPath {
 		return nil, false, nil
 	}
 
-	rootColumns := root.Relational().OutputCols
+	rootPhysicalProps := o.mem.RootProps()
+
+	if !rootPhysicalProps.Ordering.Any() {
+		return nil, false, nil
+	}
 
 	// TODO(radu): if we want to support more cases, we should use optgen to write
 	// the rules.
@@ -175,7 +188,7 @@ func (o *Optimizer) TryPlaceholderFastPath() (_ opt.Expr, ok bool, err error) {
 
 	// Success!
 	newPrivate := scan.ScanPrivate
-	newPrivate.Cols = rootColumns
+	newPrivate.Cols = rootRelProps.OutputCols
 	newPrivate.Index = foundIndex.Ordinal()
 
 	span := make(memo.ScalarListExpr, numConstrained)
@@ -198,8 +211,8 @@ func (o *Optimizer) TryPlaceholderFastPath() (_ opt.Expr, ok bool, err error) {
 		ScanPrivate: newPrivate,
 	}
 	placeholderScan = o.mem.AddPlaceholderScanToGroup(placeholderScan, root)
-	o.mem.SetBestProps(placeholderScan, rootProps, &physical.Provided{}, 1.0 /* cost */)
-	o.mem.SetRoot(placeholderScan, rootProps)
+	o.mem.SetBestProps(placeholderScan, rootPhysicalProps, &physical.Provided{}, 1.0 /* cost */)
+	o.mem.SetRoot(placeholderScan, rootPhysicalProps)
 
 	if util.CrdbTestBuild && !o.mem.IsOptimized() {
 		return nil, false, errors.AssertionFailedf("IsOptimized() should be true")

pkg/sql/opt/xform/testdata/placeholder-fast-path/scan

@@ -87,13 +87,32 @@ SELECT v+1 FROM kv WHERE k = $1
 ----
 no fast path
 
+# The fast path should not kick in because the estimated row count is too high.
+placeholder-fast-path
+SELECT a, b, c FROM abcd WHERE a=$1 AND b=$2
+----
+no fast path
+
+# Now inject statistics so that the estimated row count is small.
+exec-ddl
+ALTER TABLE abcd INJECT STATISTICS '[
+  {
+    "columns": ["a"],
+    "created_at": "2018-05-01 1:00:00.00000+00:00",
+    "row_count": 10,
+    "distinct_count": 5
+  }
+]'
+----
+
+# The fast path should now kick in.
 placeholder-fast-path
 SELECT a, b, c FROM abcd WHERE a=$1 AND b=$2
 ----
 placeholder-scan abcd@secondary
  ├── columns: a:1!null b:2!null c:3
  ├── has-placeholder
- ├── stats: [rows=109.89, distinct(1)=100, null(1)=0, distinct(2)=100, null(2)=0, distinct(1,2)=109.89, null(1,2)=0]
+ ├── stats: [rows=1.1, distinct(1)=1.1, null(1)=0, distinct(2)=1, null(2)=0]
  ├── fd: ()-->(1,2)
  └── span
       ├── $1
@@ -105,7 +124,7 @@ SELECT a, b, c FROM abcd WHERE b=$1 AND a=$2
 placeholder-scan abcd@secondary
  ├── columns: a:1!null b:2!null c:3
  ├── has-placeholder
- ├── stats: [rows=109.89, distinct(1)=100, null(1)=0, distinct(2)=100, null(2)=0, distinct(1,2)=109.89, null(1,2)=0]
+ ├── stats: [rows=1.1, distinct(1)=1.1, null(1)=0, distinct(2)=1, null(2)=0]
  ├── fd: ()-->(1,2)
  └── span
       ├── $2
@@ -118,7 +137,7 @@ SELECT a, b, c FROM abcd WHERE a=0 AND b=$1
 placeholder-scan abcd@secondary
  ├── columns: a:1!null b:2!null c:3
  ├── has-placeholder
- ├── stats: [rows=3.33, distinct(1)=1, null(1)=0, distinct(2)=3.33, null(2)=0, distinct(1,2)=3.33, null(1,2)=0]
+ ├── stats: [rows=0.666, distinct(1)=0.666, null(1)=0, distinct(2)=0.666, null(2)=0, distinct(1,2)=0.666, null(1,2)=0]
  ├── fd: ()-->(1,2)
  └── span
       ├── 0
@@ -130,7 +149,7 @@ SELECT a, b, c FROM abcd WHERE a=$1 AND b=0
 placeholder-scan abcd@secondary
  ├── columns: a:1!null b:2!null c:3
  ├── has-placeholder
- ├── stats: [rows=3.33, distinct(1)=3.33, null(1)=0, distinct(2)=1, null(2)=0, distinct(1,2)=3.33, null(1,2)=0]
+ ├── stats: [rows=3.3, distinct(1)=3.3, null(1)=0, distinct(2)=1, null(2)=0]
  ├── fd: ()-->(1,2)
  └── span
       ├── $1
@@ -143,7 +162,7 @@ SELECT a, b, c FROM abcd WHERE a=1+2 AND b=$1
 placeholder-scan abcd@secondary
  ├── columns: a:1!null b:2!null c:3
  ├── has-placeholder
- ├── stats: [rows=3.33, distinct(1)=1, null(1)=0, distinct(2)=3.33, null(2)=0, distinct(1,2)=3.33, null(1,2)=0]
+ ├── stats: [rows=0.666, distinct(1)=0.666, null(1)=0, distinct(2)=0.666, null(2)=0, distinct(1,2)=0.666, null(1,2)=0]
  ├── fd: ()-->(1,2)
  └── span
       ├── 3
@@ -155,7 +174,7 @@ SELECT a, b, c FROM abcd WHERE a=fnv32a('foo') AND b=$1
 placeholder-scan abcd@secondary
  ├── columns: a:1!null b:2!null c:3
  ├── has-placeholder
- ├── stats: [rows=3.33, distinct(1)=1, null(1)=0, distinct(2)=3.33, null(2)=0, distinct(1,2)=3.33, null(1,2)=0]
+ ├── stats: [rows=0.666, distinct(1)=0.666, null(1)=0, distinct(2)=0.666, null(2)=0, distinct(1,2)=0.666, null(1,2)=0]
  ├── fd: ()-->(1,2)
  └── span
       ├── 2851307223
@@ -203,7 +222,7 @@ SELECT a, d FROM abcd WHERE d=$1 AND c=$2
 placeholder-scan abcd@secondary
  ├── columns: a:1 d:4!null
  ├── has-placeholder
- ├── stats: [rows=109.89]
+ ├── stats: [rows=1.0989]
  ├── fd: ()-->(4)
  └── span
       ├── $1
@@ -235,6 +254,19 @@ CREATE TABLE partial1 (
 )
 ----
 
+# The fast path is conditional on having a small estimated row count. Inject
+# statistics so that we don't have to worry about this aspect in tests.
+exec-ddl
+ALTER TABLE partial1 INJECT STATISTICS '[
+  {
+    "columns": ["k"],
+    "created_at": "2018-05-01 1:00:00.00000+00:00",
+    "row_count": 10,
+    "distinct_count": 10
+  }
+]'
+----
+
 # Make sure the fast path doesn't choose the cab index, getting in the way of
 # using partial_ab (which might be the better index when the placeholder is 0).
 placeholder-fast-path
@@ -250,7 +282,7 @@ SELECT a, b FROM partial1 WHERE a = $1
 placeholder-scan partial1@pseudo_ab,partial
  ├── columns: a:2!null b:3
  ├── has-placeholder
- ├── stats: [rows=330, distinct(2)=100, null(2)=0]
+ ├── stats: [rows=3.3, distinct(2)=1, null(2)=0]
  ├── fd: ()-->(2)
  └── span
       └── $1