sql: replicating JSON empty array ordering found in Postgres

Currently, cockroachdb#97928 and cockroachdb#99275 are responsible for laying out a
lexicographical ordering for JSON columns to be forward indexable in
nature. This ordering is based on the rules posted by Postgres and is
in cockroachdb#99849.

However, Postgres currently sorts the empty JSON array before any other
JSON values. A Postgres bug report for this has been opened:
https://www.postgresql.org/message-id/17873-826fdc8bbcace4f1%40postgresql.org

This PR intends on replicating the Postgres behavior.

Fixes cockroachdb#105668

Epic: CRDB-24501

Release note: None

Author: Shivs11

docs/tech-notes/jsonb_forward_indexing.md

@@ -44,10 +44,18 @@ The following rules were kept in mind while designing this form of encoding, as
 5. Objects with an equal number of key value pairs are compared in the order:
 `key1`, `value1`, `key2`, `value2`, ….
 
+**NOTE:** There is one exception to these rules, which is neither documented by
+Postgres, nor mentioned in the source code: empty arrays are the minimum JSON
+value. As far as we can tell, this is a Postgres bug that has existed for some
+time. We've decided to replicate this behavior to remain consistent with
+Postgres. We've filed a [Postgres bug report](https://www.postgresql.org/message-id/17873-826fdc8bbcace4f1%40postgresql.org)
+to track the issue.
+
 In order to satisfy property 1 at all times, tags are defined in an increasing order of bytes. 
 These tags will also have to be defined in a way where the tag representing an object is a large byte representation 
 for a hexadecimal value (such as 0xff) and the subsequent objects have a value 1 less than the previous one,
-where the ordering is described in point 1 above. 
+where the ordering is described in point 1 above. There is a special tag for empty JSON arrays
+in order to handle the special case of empty arrays being ordered before all other JSON values.
 
 Additionally, tags representing terminators will also be defined. There will be two terminators, one for the ascending designation and the other for the descending one, and will be required to denote the end of a key encoding of the following JSON values: Objects, Arrays, Number and Strings. JSON Boolean and JSON Null are not required to have the terminator since they do not have variable length encoding due to the presence of a single tag (as explained later in this document).
 

pkg/sql/logictest/testdata/logic_test/json_index

@@ -20,13 +20,13 @@ INSERT INTO t VALUES
 query T
 SELECT x FROM t ORDER BY x
 ----
+[]
 "a"
 "aa"
 "abcdefghi"
 "b"
 1
 100
-[]
 {"a": "b"}
 
 
@@ -38,13 +38,13 @@ INSERT INTO t VALUES
 query T
 SELECT x FROM t@t_pkey ORDER BY x
 ----
+[]
 "a"
 "aa"
 "abcdefghi"
 "b"
 1
 100
-[]
 {"a": "b"}
 
 # Use the index for point lookups.
@@ -77,12 +77,12 @@ query T
 SELECT x FROM t@t_pkey WHERE x > '1' ORDER BY x
 ----
 100
-[]
 {"a": "b"}
 
 query T
 SELECT x FROM t@t_pkey WHERE x < '1' ORDER BY x
 ----
+[]
 "a"
 "aa"
 "abcdefghi"
@@ -92,12 +92,12 @@ SELECT x FROM t@t_pkey WHERE x < '1' ORDER BY x
 query T
 SELECT x FROM t@t_pkey WHERE x > '1' OR x < '1' ORDER BY x
 ----
+[]
 "a"
 "aa"
 "abcdefghi"
 "b"
 100
-[]
 {"a": "b"}
 
 query T
@@ -109,12 +109,12 @@ query T
 SELECT x FROM t@t_pkey WHERE x > '1' OR x < '1' ORDER BY x DESC
 ----
 {"a": "b"}
-[]
 100
 "b"
 "abcdefghi"
 "aa"
 "a"
+[]
 
 # Adding more primitive JSON values.
 statement ok
@@ -129,6 +129,7 @@ INSERT INTO t VALUES
 query T
 SELECT x FROM t@t_pkey ORDER BY x
 ----
+[]
 null
 "Testing Punctuation?!."
 "a"
@@ -141,18 +142,17 @@ null
 100
 false
 true
-[]
 {"a": "b"}
 
 query T
 SELECT x FROM t@t_pkey WHERE x > 'true' ORDER BY x
 ----
-[]
 {"a": "b"}
 
 query T
 SELECT x FROM t@t_pkey WHERE x < 'false' ORDER BY x
 ----
+[]
 null
 "Testing Punctuation?!."
 "a"
@@ -330,12 +330,12 @@ query T
 SELECT x FROM t@t_pkey ORDER BY x
 ----
 NULL
+[]
 null
 "crdb"
 1
 false
 true
-[]
 [1, 2, 3]
 {}
 {"a": "b", "c": "d"}
@@ -346,24 +346,24 @@ SELECT x FROM t@t_pkey ORDER BY x DESC
 {"a": "b", "c": "d"}
 {}
 [1, 2, 3]
-[]
 true
 false
 1
 "crdb"
 null
+[]
 NULL
 
 # Test to show JSON Null is different from NULL.
 query T
 SELECT x FROM t@t_pkey WHERE x IS NOT NULL ORDER BY x
 ----
+[]
 null
 "crdb"
 1
 false
 true
-[]
 [1, 2, 3]
 {}
 {"a": "b", "c": "d"}
@@ -446,12 +446,12 @@ INSERT INTO t VALUES
 query T
 SELECT x FROM t@i ORDER BY x;
 ----
+[]
 null
 "crdb"
 1
 false
 true
-[]
 [null]
 [1]
 [{"a": "b"}]

pkg/sql/opt/exec/execbuilder/testdata/json

@@ -205,7 +205,7 @@ vectorized: true
 • scan
   missing stats
   table: t@t_pkey
-  spans: [/'null' - /'null'] [/'""' - /'""'] [/'[]' - /'[]'] [/'{}' - /'{}']
+  spans: [/'[]' - /'[]'] [/'null' - /'null'] [/'""' - /'""'] [/'{}' - /'{}']
 
 # Multicolumn index, including JSONB
 
@@ -252,20 +252,20 @@ INSERT INTO composite VALUES (1, '1.00'::JSONB), (2, '1'::JSONB), (3, '2'::JSONB
  (4, '3.0'::JSONB), (5, '"a"'::JSONB)
 ----
 CPut /Table/108/1/1/0 -> /TUPLE/
-InitPut /Table/108/2/"G*\x02\x00\x00\x89\x88" -> /BYTES/0x2f0f0c200000002000000403348964
+InitPut /Table/108/2/"H*\x02\x00\x00\x89\x88" -> /BYTES/0x2f0f0c200000002000000403348964
 CPut /Table/108/1/2/0 -> /TUPLE/
-InitPut /Table/108/2/"G*\x02\x00\x00\x8a\x88" -> /BYTES/
+InitPut /Table/108/2/"H*\x02\x00\x00\x8a\x88" -> /BYTES/
 CPut /Table/108/1/3/0 -> /TUPLE/
-InitPut /Table/108/2/"G*\x04\x00\x00\x8b\x88" -> /BYTES/
+InitPut /Table/108/2/"H*\x04\x00\x00\x8b\x88" -> /BYTES/
 CPut /Table/108/1/4/0 -> /TUPLE/
-InitPut /Table/108/2/"G*\x06\x00\x00\x8c\x88" -> /BYTES/0x2f0f0c20000000200000040334891e
+InitPut /Table/108/2/"H*\x06\x00\x00\x8c\x88" -> /BYTES/0x2f0f0c20000000200000040334891e
 CPut /Table/108/1/5/0 -> /TUPLE/
-InitPut /Table/108/2/"F\x12a\x00\x01\x00\x8d\x88" -> /BYTES/
+InitPut /Table/108/2/"G\x12a\x00\x01\x00\x8d\x88" -> /BYTES/
 
 query T kvtrace
 SELECT j FROM composite where j = '1.00'::JSONB
 ----
-Scan /Table/108/2/"G*\x02\x00\x0{0"-1"}
+Scan /Table/108/2/"H*\x02\x00\x0{0"-1"}
 
 query T
 SELECT j FROM composite ORDER BY j;

pkg/sql/rowenc/keyside/json.go

@@ -79,7 +79,7 @@ func decodeJSONKey(buf []byte, dir encoding.Direction) (json.JSON, []byte, error
 		}
 		buf = buf[1:] // removing the terminator
 		jsonVal = json.FromDecimal(dec)
-	case encoding.JSONArray, encoding.JSONArrayDesc:
+	case encoding.JSONArray, encoding.JSONArrayDesc, encoding.JsonEmptyArray, encoding.JsonEmptyArrayDesc:
 		jsonVal, buf, err = decodeJSONArray(buf, dir)
 		if err != nil {
 			return nil, nil, errors.NewAssertionErrorWithWrappedErrf(err, "could not decode JSON Array")

pkg/util/encoding/encoding.go

@@ -107,13 +107,18 @@ const (
 
 	// Defining different key markers, for the ascending designation,
 	// for handling different JSON values.
-	jsonNullKeyMarker   = voidMarker + 1
-	jsonStringKeyMarker = jsonNullKeyMarker + 1
-	jsonNumberKeyMarker = jsonStringKeyMarker + 1
-	jsonFalseKeyMarker  = jsonNumberKeyMarker + 1
-	jsonTrueKeyMarker   = jsonFalseKeyMarker + 1
-	jsonArrayKeyMarker  = jsonTrueKeyMarker + 1
-	jsonObjectKeyMarker = jsonArrayKeyMarker + 1
+
+	// Postgres currently has a special case (maybe a bug) where the empty JSON
+	// Array sorts before all other JSON values. See the bug report:
+	// https://www.postgresql.org/message-id/17873-826fdc8bbcace4f1%40postgresql.org
+	jsonEmptyArrayKeyMarker = voidMarker + 1
+	jsonNullKeyMarker       = jsonEmptyArrayKeyMarker + 1
+	jsonStringKeyMarker     = jsonNullKeyMarker + 1
+	jsonNumberKeyMarker     = jsonStringKeyMarker + 1
+	jsonFalseKeyMarker      = jsonNumberKeyMarker + 1
+	jsonTrueKeyMarker       = jsonFalseKeyMarker + 1
+	jsonArrayKeyMarker      = jsonTrueKeyMarker + 1
+	jsonObjectKeyMarker     = jsonArrayKeyMarker + 1
 
 	arrayKeyTerminator           byte = 0x00
 	arrayKeyDescendingTerminator byte = 0xFF
@@ -127,13 +132,14 @@ const (
 
 	// Defining different key markers, for the descending designation,
 	// for handling different JSON values.
-	jsonNullKeyDescendingMarker   = jsonObjectKeyMarker + 7
-	jsonStringKeyDescendingMarker = jsonNullKeyDescendingMarker - 1
-	jsonNumberKeyDescendingMarker = jsonStringKeyDescendingMarker - 1
-	jsonFalseKeyDescendingMarker  = jsonNumberKeyDescendingMarker - 1
-	jsonTrueKeyDescendingMarker   = jsonFalseKeyDescendingMarker - 1
-	jsonArrayKeyDescendingMarker  = jsonTrueKeyDescendingMarker - 1
-	jsonObjectKeyDescendingMarker = jsonArrayKeyDescendingMarker - 1
+	jsonEmptyArrayKeyDescendingMarker = jsonObjectKeyMarker + 8
+	jsonNullKeyDescendingMarker       = jsonEmptyArrayKeyDescendingMarker - 1
+	jsonStringKeyDescendingMarker     = jsonNullKeyDescendingMarker - 1
+	jsonNumberKeyDescendingMarker     = jsonStringKeyDescendingMarker - 1
+	jsonFalseKeyDescendingMarker      = jsonNumberKeyDescendingMarker - 1
+	jsonTrueKeyDescendingMarker       = jsonFalseKeyDescendingMarker - 1
+	jsonArrayKeyDescendingMarker      = jsonTrueKeyDescendingMarker - 1
+	jsonObjectKeyDescendingMarker     = jsonArrayKeyDescendingMarker - 1
 
 	// Terminators for JSON Key encoding.
 	jsonKeyTerminator           byte = 0x00
@@ -1789,6 +1795,9 @@ const (
 	JSONArrayDesc  Type = 39
 	JSONObject     Type = 40
 	JSONObjectDesc Type = 41
+	// Special case
+	JsonEmptyArray     Type = 42
+	JsonEmptyArrayDesc Type = 43
 )
 
 // typMap maps an encoded type byte to a decoded Type. It's got 256 slots, one
@@ -1849,6 +1858,10 @@ func slowPeekType(b []byte) Type {
 			return JSONArray
 		case m == jsonArrayKeyDescendingMarker:
 			return JSONArrayDesc
+		case m == jsonEmptyArrayKeyMarker:
+			return JsonEmptyArray
+		case m == jsonEmptyArrayKeyDescendingMarker:
+			return JsonEmptyArrayDesc
 		case m == jsonObjectKeyMarker:
 			return JSONObject
 		case m == jsonObjectKeyDescendingMarker:
@@ -2009,10 +2022,12 @@ func PeekLength(b []byte) (int, error) {
 		length, err := getArrayOrJSONLength(b[1:], dir, IsJSONKeyDone)
 		return 1 + length, err
 	case jsonArrayKeyMarker, jsonArrayKeyDescendingMarker,
-		jsonObjectKeyMarker, jsonObjectKeyDescendingMarker:
+		jsonObjectKeyMarker, jsonObjectKeyDescendingMarker,
+		jsonEmptyArrayKeyMarker, jsonEmptyArrayKeyDescendingMarker:
 		dir := Ascending
 		if (m == jsonArrayKeyDescendingMarker) ||
-			(m == jsonObjectKeyDescendingMarker) {
+			(m == jsonObjectKeyDescendingMarker) ||
+			(m == jsonEmptyArrayKeyDescendingMarker) {
 			dir = Descending
 		}
 		// removing the starter tag
@@ -3500,11 +3515,17 @@ func EncodeJSONTrueKeyMarker(buf []byte, dir Direction) []byte {
 
 // EncodeJSONArrayKeyMarker adds a JSON Array key encoding marker
 // to buf and returns the new buffer.
-func EncodeJSONArrayKeyMarker(buf []byte, dir Direction) []byte {
+func EncodeJSONArrayKeyMarker(buf []byte, dir Direction, arrayLength int64) []byte {
 	switch dir {
 	case Ascending:
+		if arrayLength == 0 {
+			return append(buf, jsonEmptyArrayKeyMarker)
+		}
 		return append(buf, jsonArrayKeyMarker)
 	case Descending:
+		if arrayLength == 0 {
+			return append(buf, jsonEmptyArrayKeyDescendingMarker)
+		}
 		return append(buf, jsonArrayKeyDescendingMarker)
 	default:
 		panic("invalid direction")
@@ -3621,15 +3642,15 @@ func ValidateAndConsumeJSONKeyMarker(buf []byte, dir Direction) ([]byte, Type, e
 	case Descending:
 		switch typ {
 		case JSONNullDesc, JSONNumberDesc, JSONStringDesc, JSONFalseDesc,
-			JSONTrueDesc, JSONArrayDesc, JSONObjectDesc:
+			JSONTrueDesc, JSONArrayDesc, JSONObjectDesc, JsonEmptyArrayDesc:
 			return buf[1:], typ, nil
 		default:
 			return nil, Unknown, errors.Newf("invalid type found %s", typ)
 		}
 	case Ascending:
 		switch typ {
 		case JSONNull, JSONNumber, JSONString, JSONFalse, JSONTrue, JSONArray,
-			JSONObject:
+			JSONObject, JsonEmptyArray:
 			return buf[1:], typ, nil
 		default:
 			return nil, Unknown, errors.Newf("invalid type found %s", typ)

pkg/util/encoding/type_string.go

@@ -606,10 +606,20 @@ func (j *jsonEncoded) AreKeysSorted() bool {
 	return decoded.AreKeysSorted()
 }
 
-func (j *jsonEncoded) Compare(other JSON) (int, error) {
+func (j *jsonEncoded) Compare(other JSON) (_ int, err error) {
 	if other == nil {
 		return -1, nil
 	}
+	// We must first check for the special case of empty arrays, which are the
+	// minimum JSON value.
+	switch {
+	case isEmptyArray(j) && isEmptyArray(other):
+		return 0, nil
+	case isEmptyArray(j):
+		return -1, nil
+	case isEmptyArray(other):
+		return 1, nil
+	}
 	if cmp := cmpJSONTypes(j.Type(), other.Type()); cmp != 0 {
 		return cmp, nil
 	}