sql: enabling forward indexes and ORDERBY on JSONB columns

Currently, cockroachdb#97928 outlines the scheme for JSONB encoding
and decoding for forward indexes. However, the PR doesn't
enable this feature to our users. This current PR aims
to allow forward indexes on JSONB columns. The presence
of a lexicographical ordering, as described in cockroachdb#97928,
shall now allow primary and secondary indexes on JSONB
columns along with the ability to use ORDER BY filter
in their queries.

Additionally, JSON values consist of decimal numbers
and containers, such as Arrays and Objects, which can
contain these decimal numbers. In order to preserve
the values after the decimal, JSONB columns are now
required to be composite in nature. This shall enable
such values to be stored in both the key and the value
side of a K/V pair in hopes of receiving the exact value.

Fixes: cockroachdb#35706

Release note (sql change): This PR adds support for enabling
forward indexes and ordering on JSON values.

Epic: CRDB-24501

pkg/sql/catalog/colinfo/col_type_info.go

@@ -146,7 +146,7 @@ func ColumnTypeIsIndexable(t *types.T) bool {
 // using an inverted index.
 func ColumnTypeIsInvertedIndexable(t *types.T) bool {
 	switch t.Family() {
-	case types.ArrayFamily, types.StringFamily:
+	case types.JsonFamily, types.ArrayFamily, types.StringFamily:
 		return true
 	}
 	return ColumnTypeIsOnlyInvertedIndexable(t)
@@ -162,7 +162,6 @@ func ColumnTypeIsOnlyInvertedIndexable(t *types.T) bool {
 		t = t.ArrayContents()
 	}
 	switch t.Family() {
-	case types.JsonFamily:
 	case types.GeographyFamily:
 	case types.GeometryFamily:
 	case types.TSVectorFamily:
@@ -183,7 +182,7 @@ func MustBeValueEncoded(semanticType *types.T) bool {
 		default:
 			return MustBeValueEncoded(semanticType.ArrayContents())
 		}
-	case types.JsonFamily, types.TupleFamily, types.GeographyFamily, types.GeometryFamily:
+	case types.TupleFamily, types.GeographyFamily, types.GeometryFamily:
 		return true
 	case types.TSVectorFamily, types.TSQueryFamily:
 		return true

pkg/sql/catalog/colinfo/column_type_properties.go

@@ -52,6 +52,7 @@ func CanHaveCompositeKeyEncoding(typ *types.T) bool {
 	switch typ.Family() {
 	case types.FloatFamily,
 		types.DecimalFamily,
+		types.JsonFamily,
 		types.CollatedStringFamily:
 		return true
 	case types.ArrayFamily:
@@ -75,7 +76,6 @@ func CanHaveCompositeKeyEncoding(typ *types.T) bool {
 		types.UuidFamily,
 		types.INetFamily,
 		types.TimeFamily,
-		types.JsonFamily,
 		types.TimeTZFamily,
 		types.BitFamily,
 		types.GeometryFamily,

pkg/sql/catalog/colinfo/column_type_properties_test.go

@@ -52,7 +52,7 @@ func TestCanHaveCompositeKeyEncoding(t *testing.T) {
 		{types.IntArray, false},
 		{types.Interval, false},
 		{types.IntervalArray, false},
-		{types.Jsonb, false},
+		{types.Jsonb, true},
 		{types.Name, false},
 		{types.Oid, false},
 		{types.String, false},

pkg/sql/catalog/table_col_set.go

@@ -56,6 +56,9 @@ func (s TableColSet) ForEach(f func(col descpb.ColumnID)) {
 	s.set.ForEach(func(i int) { f(descpb.ColumnID(i)) })
 }
 
+// Copy returns a copy of s which can be modified independently.
+func (s TableColSet) Copy() TableColSet { return TableColSet{set: s.set.Copy()} }
+
 // SubsetOf returns true if s is a subset of other.
 func (s TableColSet) SubsetOf(other TableColSet) bool {
 	return s.set.SubsetOf(other.set)

pkg/sql/catalog/tabledesc/structured.go

@@ -714,6 +714,14 @@ func (desc *Mutable) allocateIndexIDs(columnNames map[string]descpb.ColumnID) er
 			colIDs = idx.CollectKeyColumnIDs()
 		}
 
+		// Inverted indexes don't store composite values in the individual
+		// paths present. The composite values will be differentiated in
+		// the primary index itself.
+		compositeColIDsLocal := compositeColIDs.Copy()
+		if isInverted {
+			compositeColIDsLocal.Remove(invID)
+		}
+
 		// StoreColumnIDs are derived from StoreColumnNames just like KeyColumnIDs
 		// derives from KeyColumnNames.
 		// For primary indexes this set of columns is typically defined as the set
@@ -755,12 +763,12 @@ func (desc *Mutable) allocateIndexIDs(columnNames map[string]descpb.ColumnID) er
 		// or in the primary key whose type has a composite encoding, like DECIMAL
 		// for instance.
 		for _, colID := range idx.IndexDesc().KeyColumnIDs {
-			if compositeColIDs.Contains(colID) {
+			if compositeColIDsLocal.Contains(colID) {
 				idx.IndexDesc().CompositeColumnIDs = append(idx.IndexDesc().CompositeColumnIDs, colID)
 			}
 		}
 		for _, colID := range idx.IndexDesc().KeySuffixColumnIDs {
-			if compositeColIDs.Contains(colID) {
+			if compositeColIDsLocal.Contains(colID) {
 				idx.IndexDesc().CompositeColumnIDs = append(idx.IndexDesc().CompositeColumnIDs, colID)
 			}
 		}

pkg/sql/colenc/encode.go

@@ -707,6 +707,9 @@ func isComposite(vec coldata.Vec, row int) bool {
 	case types.DecimalFamily:
 		d := tree.DDecimal{Decimal: vec.Decimal()[row]}
 		return d.IsComposite()
+	case types.JsonFamily:
+		j := tree.DJSON{JSON: vec.JSON().Get(row)}
+		return j.IsComposite()
 	default:
 		d := vec.Datum().Get(row)
 		if cdatum, ok := d.(tree.CompositeDatum); ok {

pkg/sql/colenc/key.go

@@ -180,6 +180,21 @@ func encodeKeys[T []byte | roachpb.Key](
 			}
 			kys[r] = b
 		}
+	case types.JsonFamily:
+		jsonVector := vec.JSON()
+		for r := 0; r < count; r++ {
+			b := kys[r]
+			if partialIndexAndNullCheck(kys, r, start, nulls, dir) {
+				continue
+			}
+			var err error
+			jsonObj := jsonVector.Get(r + start)
+			b, err = jsonObj.EncodeForwardIndex(b, dir)
+			if err != nil {
+				return err
+			}
+			kys[r] = b
+		}
 	default:
 		if buildutil.CrdbTestBuild {
 			if typeconv.TypeFamilyToCanonicalTypeFamily(typ.Family()) != typeconv.DatumVecCanonicalTypeFamily {

pkg/sql/colencoding/key_encoding.go

@@ -187,12 +187,21 @@ func decodeTableKeyToCol(
 		}
 		vecs.IntervalCols[colIdx][rowIdx] = d
 	case types.JsonFamily:
-		// Don't attempt to decode the JSON value. Instead, just return the
-		// remaining bytes of the key.
-		var jsonLen int
-		jsonLen, err = encoding.PeekLength(key)
-		vecs.JSONCols[colIdx].Bytes.Set(rowIdx, key[:jsonLen])
-		rkey = key[jsonLen:]
+		// Decode the JSON, and then store the bytes in the
+		// vector in the value-encoded format.
+		// TODO (shivam): Make it possible for the vector to store
+		// key-encoded JSONs instead of value-encoded JSONs.
+		var d tree.Datum
+		encDir := encoding.Ascending
+		if dir == catenumpb.IndexColumn_DESC {
+			encDir = encoding.Descending
+		}
+		d, rkey, err = keyside.Decode(da, valType, key, encDir)
+		json, ok := d.(*tree.DJSON)
+		if !ok {
+			return nil, false, scratch, errors.AssertionFailedf("Could not type assert into DJSON")
+		}
+		vecs.JSONCols[colIdx].Set(rowIdx, json.JSON)
 	case types.EncodedKeyFamily:
 		// Don't attempt to decode the inverted key.
 		keyLen, err := encoding.PeekLength(key)

pkg/sql/logictest/testdata/logic_test/distsql_stats

@@ -1257,7 +1257,7 @@ ORDER BY
 statistics_name  column_names  row_count  null_count  has_histogram
 s                {a}           3          0           true
 s                {b}           3          0           true
-s                {j}           3          0           false
+s                {j}           3          0           true
 s                {rowid}       3          0           true
 
 # Test that non-index columns have histograms collected for them, with
@@ -2348,7 +2348,7 @@ SHOW STATISTICS USING JSON FOR TABLE j;
 statement ok
 ALTER TABLE j INJECT STATISTICS '$j_stats'
 
-statement error pq: cannot create partial statistics on an inverted index column
+statement error pq: table j does not contain a non-partial forward index with j as a prefix column
 CREATE STATISTICS j_partial ON j FROM j USING EXTREMES;
 
 statement ok