ganeshashree commented on code in PR #58034: URL: https://github.com/apache/spark/pull/58034#discussion_r4282661110
########## sql/core/src/test/scala/org/apache/spark/sql/JsonObjectSuite.scala: ########## @@ -0,0 +1,1368 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.sql + +import org.antlr.v4.runtime.{CharStreams, CommonTokenStream} +import org.antlr.v4.runtime.atn.PredictionMode + +import org.apache.spark.SparkRuntimeException +import org.apache.spark.sql.catalyst.analysis.Star +import org.apache.spark.sql.catalyst.analysis.TypeCheckResult.DataTypeMismatch +import org.apache.spark.sql.catalyst.expressions.{Cast, Collate, Expression, JsonConstructorNullBehavior, JsonImplicitFormatCarrier, JsonObjectExpr, Literal} +import org.apache.spark.sql.catalyst.parser.{ParseException, SparkParserBailErrorStrategy, SqlBaseLexer, SqlBaseParser} +import org.apache.spark.sql.catalyst.plans.logical.{Filter, Join} +import org.apache.spark.sql.execution.{ProjectExec, WholeStageCodegenExec} +import org.apache.spark.sql.internal.SQLConf +import org.apache.spark.sql.test.SharedSparkSession +import org.apache.spark.sql.types.{ArrayType, CharType, GeographyType, GeometryType, IntegerType, MapType, StringType, StructType, VarcharType} + +/** + * End-to-end tests for the SQL:2016 `JSON_OBJECT` constructor function. + */ +class JsonObjectSuite extends QueryTest with SharedSparkSession { + import testImplicits._ + + test("in the comma form, colon-bearing arguments bind as colon members (documented precedence)") { + // `JSON_OBJECT(a:b, c:d)` matches both list forms: the colon-member list and the comma + // (MySQL-compat) list. The colon-member alternative is declared first, so it wins -- `a:b, c:d` + // reads as the two members (a -> b) and (c -> d), the SQL-standard `key : value` reading, not + // as a single comma-form pair whose key and value are `:`-extractions. + checkAnswer( + sql("SELECT json_object(a:b, c:d) FROM VALUES ('k1','v1','k2','v2') t(a,b,c,d)"), + Row("""{"k1":"v1","k2":"v2"}""")) + // Parenthesizing each argument forces the comma form (one member whose key/value are the + // `:`-extractions). The non-VARIANT extraction target then errors -- proving it parsed as one + // comma-form pair, not two colon members (which resolve cleanly above). + val e = intercept[AnalysisException] { + sql("SELECT json_object((a:b), (c:d)) FROM VALUES ('k1','v1','k2','v2') t(a,b,c,d)").collect() + } + assert(e.getCondition == "COLUMN_IS_NOT_VARIANT_TYPE") + } + + test("basic object from key-value pairs using VALUE keyword") { + checkAnswer( + sql("SELECT json_object('id' VALUE 7, 'name' VALUE 'Ada')"), + Row("""{"id":7,"name":"Ada"}""")) + } + + test("construct object using optional KEY keyword") { + checkAnswer( + sql("SELECT json_object(KEY 'id' VALUE 7, KEY 'name' VALUE 'Ada')"), + Row("""{"id":7,"name":"Ada"}""")) + } + + test("construct object using colon syntax") { + checkAnswer( + sql("SELECT json_object('id': 7, 'name': 'Ada')"), + Row("""{"id":7,"name":"Ada"}""")) + } + + test("colon member with a column-valued value") { + // The `key : value` member overlaps the semi-structured extraction form `expr : path` + // (`primaryExpression COLON semiStructuredExtractionPath`), which also accepts a bare + // identifier on the right. Confirm a column-valued colon member still parses as a JSON_OBJECT + // pair rather than as `'k' : v` semi-structured extraction. + checkAnswer( + sql("SELECT json_object('k': v) FROM VALUES ('x') t(v)"), + Row("""{"k":"x"}""")) + } + + test("a colon member splits at the first top-level colon after the key") { + val from = + """FROM (SELECT parse_json('{"x":1,"y":"s"}') AS v, 'kk' AS k, 'e' AS `end`, 'c' AS `case`)""" + Seq( + "json_object('k': v:x)" -> """{"k":1}""", + "json_object(k: v:x)" -> """{"kk":1}""", + "json_object('k': v:x ABSENT ON NULL)" -> """{"k":1}""", + "json_object(k: v:x RETURNING STRING)" -> """{"kk":1}""", + "json_object('a': v:x, k: v:y)" -> """{"a":1,"kk":"s"}""", + "json_object((v:y::string) : 1)" -> """{"s":1}""", + "json_object(v:y::string VALUE v:x)" -> """{"s":1}""", + "json_object(CASE WHEN v:x::int = 1 THEN 'a' END : v:y)" -> """{"a":"s"}""", + // `end` and `case` are non-reserved and usable as column names: the separator check must read + // each CASE/END token as keyword or column by its neighbour, not count it blindly. + "json_object(CASE WHEN end = 'e' AND v:x::int = 1 THEN 'a' END : v:y)" -> """{"a":"s"}""", + "json_object(case : v:x)" -> """{"c":1}""", + "json_object(array('a')[0] : v:x)" -> """{"a":1}""", + "json_object(end : v:x)" -> """{"e":1}""" + ).foreach { case (call, expected) => + checkAnswer(sql(s"SELECT $call $from"), Row(expected)) + } + checkAnswer( + spark.sql(s"SELECT json_object('k': v:x) $from", Map("v" -> "zzz")), + Row("""{"k":1}""")) + checkAnswer( + spark.sql("SELECT json_object('k' :p VALUE 1), 'a' :p", Map("p" -> "x")), + Row("""{"kx":1}""", "ax")) + } + + test("a case-named column with a postfix operator is a column, not a CASE keyword") { + // `case` is non-reserved, so `case[0]` (subscript) and `case COLLATE ...` apply a postfix + // operator to a column named `case`; neither `CASE [` nor `CASE COLLATE` begins a real CASE + // expression (`[` has no array-literal form, COLLATE needs a left operand). The separator check + // must read the token right after `case` as a postfix on an operand -- like the already-handled + // `.` -- so the top-level colon is still the member separator, not swallowed as a CASE branch. + checkAnswer( + sql("""SELECT json_object(case[0] : v:x) + |FROM (SELECT array('c') AS `case`, parse_json('{"x":1}') AS v)""".stripMargin), + Row("""{"c":1}""")) + checkAnswer( + sql("""SELECT json_object(case COLLATE UTF8_LCASE : v:x) + |FROM (SELECT 'c' AS `case`, parse_json('{"x":1}') AS v)""".stripMargin), + Row("""{"c":1}""")) + } + + test("a colon key cannot coalesce a trailing named parameter marker") { + // `'k' :p` coalesces the marker into the key under the VALUE form, but in the colon form the + // first top-level `:` is the member separator, so a bare marker-coalesced key is a clean parse + // error (never a silent mis-split). Parenthesize the key or use the VALUE form. + val e = intercept[ParseException] { + spark.sql("SELECT json_object('k' :p : 1)", Map("p" -> "x")) + } + assert(e.getCondition == "PARSE_SYNTAX_ERROR") + checkAnswer( + spark.sql("SELECT json_object(('k' :p) : 1)", Map("p" -> "x")), Row("""{"kx":1}""")) + checkAnswer( + spark.sql("SELECT json_object('k' :p VALUE 1)", Map("p" -> "x")), Row("""{"kx":1}""")) + } + + test("a simple CASE over a named-parameter operand is a valid colon key") { + // `CASE :p WHEN ... END : value`: the `:p` operand must parse as a named-parameter marker, not + // be read as the member separator. `CASE :` is ambiguous with a `case` column before the + // separator (`case : value`), so the separator check decides it by looking ahead for the + // CASE's WHEN -- present here, so `CASE` is the keyword and `:p` stays its operand. + val from = "FROM (SELECT parse_json('{\"x\":1}') AS v)" + checkAnswer( + spark.sql( + s"SELECT json_object(CASE :p WHEN 1 THEN 'one' ELSE 'other' END : v:x) $from", + Map("p" -> 1)), + Row("""{"one":1}""")) + checkAnswer( + spark.sql( + s"SELECT json_object(CASE :p WHEN 1 THEN 'one' ELSE 'other' END : v:x) $from", + Map("p" -> 2)), + Row("""{"other":1}""")) + } + + test("a CASE-valued value does not flip a bare `case` key to the CASE keyword") { + // `case : CASE WHEN ... END`: the value's WHEN belongs to the value's CASE, not to the key, so + // the WHEN lookahead must stop at that inner CASE and keep the leading `case` a column with the + // top-level colon as the member separator. Guards the lookahead against over-triggering. + checkAnswer( + sql("""SELECT json_object(case : CASE WHEN v:x::int = 1 THEN 'y' ELSE 'z' END) + |FROM (SELECT 'c' AS `case`, parse_json('{"x":1}') AS v)""".stripMargin), + Row("""{"c":"y"}""")) + } + + test("a `case` column inside a simple-CASE colon key is a column, not a nested CASE") { + // The simple-CASE key's own body references a column named `case` (the THEN result). The + // key-colon lookahead must not count that `case` as a nested CASE...END, or it would consume + // the real END and misread the `:p` operand marker as the member separator. + withSQLConf(SQLConf.ANSI_ENABLED.key -> "false") { + val from = "FROM (SELECT 'c' AS `case`, 'v' AS value)" + checkAnswer( + spark.sql(s"SELECT json_object(CASE :p WHEN 1 THEN case ELSE 'x' END : value) $from", + Map("p" -> 1)), + Row("""{"c":"v"}""")) + checkAnswer( + spark.sql(s"SELECT json_object(CASE :p WHEN 1 THEN case ELSE 'x' END : value) $from", + Map("p" -> 2)), + Row("""{"x":"v"}""")) + // A postfix operator on the body `case` column is the case most likely to be miscounted as a + // nested CASE...END by the key-colon lookahead. + val q = s"SELECT json_object(CASE :p WHEN 1 THEN case::string ELSE 'x' END : value) $from" + checkAnswer(spark.sql(q, Map("p" -> 1)), Row("""{"c":"v"}""")) + } + } + + test("a non-foldable key with escapable characters is JSON-escaped") { + // A dynamic (non-foldable) key is written through the generator's field-name path, which must + // escape quotes and backslashes exactly as a quoted key would. + val df = Seq(("key\"with\\chars", "v")).toDF("k", "v") + checkAnswer( + df.selectExpr("json_object(k VALUE v)"), + Row("""{"key\"with\\chars":"v"}""")) + } + + test("a bare `when` column value does not flip a `case` colon key to the CASE keyword") { + // Outside ANSI mode `when` is a bare column; its WHEN token must not be read as the WHEN of a + // CASE keyword in the `case` key. A lone `when` value has no closing END, so `case` stays a + // column and the colon stays the member separator. + withSQLConf(SQLConf.ANSI_ENABLED.key -> "false") { + checkAnswer( + sql("""SELECT json_object(case : when) + |FROM (SELECT 'c' AS `case`, 'w' AS `when`)""".stripMargin), + Row("""{"c":"w"}""")) + } + } + + test("a case column nested in a colon key is a column, not a CASE keyword") { + // A `case` column in a CAST or parens is below the key's top level, where the token after it + // (`AS`, `+`) does not prove it is a column; without the depth guard the spurious CASE count + // hid the real separator colon (the parens already balance any colon inside). + checkAnswer( + sql("""SELECT json_object(CAST(case AS STRING) : v:x) + |FROM (SELECT 'c' AS `case`, parse_json('{"x":1}') AS v)""".stripMargin), + Row("""{"c":1}""")) + checkAnswer( + sql("""SELECT json_object((case + 1)::string : v:x) + |FROM (SELECT 1 AS `case`, parse_json('{"x":1}') AS v)""".stripMargin), + Row("""{"2":1}""")) + } + + test("a member whose key and value are named key and value splits at VALUE") { + checkAnswer( + sql("SELECT json_object(key VALUE value + 1) FROM VALUES ('a', 1) t(key, value)"), + Row("""{"a":2}""")) + checkAnswer( + sql("SELECT json_object(KEY key VALUE value) FROM VALUES ('a', 1) t(key, value)"), + Row("""{"a":1}""")) + checkAnswer( + sql("SELECT json_object(KEY value VALUE key) FROM VALUES ('a', 'b') t(key, value)"), + Row("""{"b":"a"}""")) + checkAnswer( + sql("SELECT json_object('x' VALUE 0, key VALUE value + 1) " + + "FROM VALUES ('a', 1) t(key, value)"), + Row("""{"x":0,"a":2}""")) + } + + test("a key column continued by an operator is the key, not the optional KEY keyword") { + // `key` is non-reserved; `key + 1` continues the `key` column into the key expression rather + // than reading `key` as the optional KEY keyword with key `+ 1`. The arithmetic key is itself a + // type error (JSON object keys must be STRING), but the report names `(key + 1)` -- proof the + // `key` column is bound into the key, not dropped. Were it read as KEY, the key would be `+ 1`. + val ex = intercept[AnalysisException] { + sql("SELECT json_object(key + 1 VALUE value) FROM VALUES (10, 'x') t(key, value)") + } + assert(ex.getCondition == "DATATYPE_MISMATCH.UNEXPECTED_INPUT_TYPE") + assert(ex.getMessage.contains("(key + 1)")) + } + + test("JSON_OBJECT members parse in SLL mode without the LL fallback") { + Seq( + "SELECT JSON_OBJECT('k': V)", + "SELECT JSON_OBJECT(K: V)", + "SELECT JSON_OBJECT('k': V:X, K: V:Y)", + "SELECT JSON_OBJECT(KEY 'id' VALUE 7)", + "SELECT JSON_OBJECT(KEY VALUE VALUE 1)", + // Exercises the optional-KEY `jsonOptionalKeyFollowsOperand` guard (`key + 1` as the key). + "SELECT JSON_OBJECT(key + 1 VALUE value)").foreach { query => + val parser = new SqlBaseParser( + new CommonTokenStream(new SqlBaseLexer(CharStreams.fromString(query)))) + parser.setErrorHandler(new SparkParserBailErrorStrategy()) + parser.getInterpreter.setPredictionMode(PredictionMode.SLL) + parser.singleStatement() + } + } + + test("construct object using comma-separated key-value syntax") { + checkAnswer( + sql("SELECT json_object('id', 7, 'name', 'Ada')"), + Row("""{"id":7,"name":"Ada"}""")) + } + + test("an odd number of arguments in the comma syntax is rejected") { + // The comma form requires paired key/value arguments; a dangling key ('name') has no value. + // JSON_OBJECT is a non-reserved keyword, so when the constructor grammar cannot match, the call + // parses as an ordinary function call and routes to the registered built-in, whose builder + // rejects the odd argument count rather than silently dropping the dangling key. + val e = intercept[AnalysisException] { + sql("SELECT json_object('id', 7, 'name')") + } + assert(e.getCondition == "WRONG_NUM_ARGS.WITHOUT_SUGGESTION") + } + + test("mixing the VALUE/colon form and the comma form is a parse error") { + // The two member-list styles are mutually exclusive grammar alternatives, so a single + // constructor cannot mix `key VALUE value` (or `key : value`) members with `key, value` ones. + Seq( + "SELECT json_object('a', 1, 'b' VALUE 2)", + "SELECT json_object('a' VALUE 1, 'b', 2)", + "SELECT json_object('a' : 1, 'b', 2)").foreach { query => + intercept[ParseException](sql(query)) + } + } + + test("construct object with NULL values (default NULL ON NULL)") { + checkAnswer( + sql("SELECT json_object('id': 7, 'v': NULL)"), + Row("""{"id":7,"v":null}""")) + } + + test("construct object with explicit NULL ON NULL") { + checkAnswer( + sql("SELECT json_object('id', 7, 'v', NULL NULL ON NULL)"), + Row("""{"id":7,"v":null}""")) + } + + test("construct object with NULL values and ABSENT ON NULL") { + checkAnswer( + sql("SELECT json_object('id': 7, 'v': NULL ABSENT ON NULL)"), + Row("""{"id":7}""")) + } + + test("construct empty object") { + checkAnswer( + sql("SELECT json_object()"), + Row("{}")) + } + + test("construct object with mixed scalar types") { + checkAnswer( + sql("""SELECT json_object('int': 42, 'str': 'hello', 'bool': true, + 'float': 3.14)"""), + Row("""{"int":42,"str":"hello","bool":true,"float":3.14}""")) + } + + test("construct object with decimal type via Jackson") { + checkAnswer( + sql("""SELECT json_object('d' VALUE CAST('123.45' AS DECIMAL(5,2)))"""), + Row("""{"d":123.45}""")) + } + + test("construct object with DATE type via Jackson") { + checkAnswer( + sql("""SELECT json_object('d' VALUE DATE'2020-01-02')"""), + Row("""{"d":"2020-01-02"}""")) + } + + test("construct object with TIMESTAMP type via Jackson") { + // Note: Jackson includes timezone offset when session timezone is set + checkAnswer( + sql("""SELECT json_object('ts' VALUE TIMESTAMP'2020-01-02 10:30:00')"""), + Row("""{"ts":"2020-01-02T10:30:00.000-08:00"}""")) + } + + test("TIMESTAMP rendering follows the session time zone on direct and routed paths") { + withSQLConf(SQLConf.SESSION_LOCAL_TIMEZONE.key -> "UTC") { + val ts = "TIMESTAMP'2020-01-02 10:30:00'" + val expected = sql(s"SELECT to_json(named_struct('ts', $ts))") + checkAnswer(sql(s"SELECT json_object('ts' VALUE $ts)"), expected) + checkAnswer(sql(s"SELECT json_object('ts', $ts)"), expected) + checkAnswer(expected, Row("""{"ts":"2020-01-02T10:30:00.000Z"}""")) + } + } + + test("struct value renders like to_json") { + // A struct value must render exactly like `to_json` of the equivalent member. + checkAnswer( + sql("SELECT json_object('s' VALUE named_struct('a', 1, 'b', 'x'))"), + Row("""{"s":{"a":1,"b":"x"}}""")) + checkAnswer( + sql("SELECT json_object('s' VALUE named_struct('a', 1, 'b', 'x'))"), + sql("SELECT to_json(named_struct('s', named_struct('a', 1, 'b', 'x')))")) + } + + test("array value renders like to_json") { + checkAnswer( + sql("SELECT json_object('a' VALUE array(1, 2, 3))"), + Row("""{"a":[1,2,3]}""")) + checkAnswer( + sql("SELECT json_object('a' VALUE array(1, 2, 3))"), + sql("SELECT to_json(named_struct('a', array(1, 2, 3)))")) + } + + test("map value renders like to_json") { + checkAnswer( + sql("SELECT json_object('m' VALUE map('x', 1, 'y', 2))"), + Row("""{"m":{"x":1,"y":2}}""")) + checkAnswer( + sql("SELECT json_object('m' VALUE map('x', 1, 'y', 2))"), + sql("SELECT to_json(named_struct('m', map('x', 1, 'y', 2)))")) + } + + test("nested complex value combining struct, array and map renders like to_json") { + val value = "named_struct('arr', array(1, 2), 'm', map('k', named_struct('n', 3)))" + checkAnswer( + sql(s"SELECT json_object('c' VALUE $value)"), + sql(s"SELECT to_json(named_struct('c', $value))")) + } + + test("struct value honors spark.sql.jsonGenerator.ignoreNullFields like to_json") { + // `ON NULL` controls only top-level members; a null field *inside* a struct value follows + // spark.sql.jsonGenerator.ignoreNullFields, like `to_json`. + val value = "named_struct('a', 1, 'b', CAST(NULL AS INT))" + Seq("true", "false").foreach { ignore => + withSQLConf(SQLConf.JSON_GENERATOR_IGNORE_NULL_FIELDS.key -> ignore) { + checkAnswer( + sql(s"SELECT json_object('s' VALUE $value)"), + sql(s"SELECT to_json(named_struct('s', $value))")) + } + } + withSQLConf(SQLConf.JSON_GENERATOR_IGNORE_NULL_FIELDS.key -> "false") { + checkAnswer(sql(s"SELECT json_object('s' VALUE $value)"), Row("""{"s":{"a":1,"b":null}}""")) + } + withSQLConf(SQLConf.JSON_GENERATOR_IGNORE_NULL_FIELDS.key -> "true") { + checkAnswer(sql(s"SELECT json_object('s' VALUE $value)"), Row("""{"s":{"a":1}}""")) + } + } + + test("top-level ON NULL and struct-internal ignoreNullFields are independent") { + // With NULL ON NULL (default) and ignoreNullFields=true, a top-level NULL member is kept as + // `null` while a null field inside a struct value is dropped. + withSQLConf(SQLConf.JSON_GENERATOR_IGNORE_NULL_FIELDS.key -> "true") { + checkAnswer( + sql("""SELECT json_object('top' VALUE CAST(NULL AS INT), + 's' VALUE named_struct('a', 1, 'b', CAST(NULL AS INT)))"""), + Row("""{"top":null,"s":{"a":1}}""")) + } + } + + test("string escaping in keys") { + checkAnswer( + sql("""SELECT json_object('key"with"quotes' VALUE 1)"""), + Row("""{"key\"with\"quotes":1}""")) + } + + // For scalar string values JSON_OBJECT must escape exactly like to_json of the equivalent + // struct (both go through the same Jackson generator); assert that equivalence rather than + // hand-encoding the escaping, which is easy to get wrong across Scala/SQL/JSON layers. + test("string escaping in values matches to_json") { + checkAnswer( + sql("""SELECT json_object('msg' VALUE 'hello +world')"""), + sql("""SELECT to_json(named_struct('msg', 'hello +world'))""")) + } + + test("string escaping with backslash matches to_json") { + checkAnswer( + sql("""SELECT json_object('path' VALUE 'c:\windows')"""), + sql("""SELECT to_json(named_struct('path', 'c:\windows'))""")) + } + + test("nested JSON_OBJECT spliced raw") { + checkAnswer( + sql("""SELECT json_object('a' VALUE json_object('b' VALUE 1))"""), + Row("""{"a":{"b":1}}""")) + checkAnswer( + sql("""SELECT json_object('a', json_object('b', 1))"""), + Row("""{"a":{"b":1}}""")) + } + + test("nested JSON_OBJECT with multiple levels") { + checkAnswer( + sql("""SELECT json_object('outer' VALUE + json_object('inner' VALUE 42, 'name' VALUE 'test'))"""), + Row("""{"outer":{"inner":42,"name":"test"}}""")) + } + + test("a nested JSON_ARRAY value is spliced raw") { + checkAnswer( + sql("SELECT json_object('a' VALUE json_array(1, 2))"), + Row("""{"a":[1,2]}""")) + } + + test("JSON_OBJECT nested directly in JSON_ARRAY is spliced as an object element") { + // The inverse nesting direction: a JSON_OBJECT in a JSON_ARRAY element position stays on the + // direct grammar path (JsonArrayValueContext), so it is spliced as a JSON object rather than + // routed through resolution and emitted as a quoted string. + checkAnswer( + sql("SELECT json_array(json_object('a', 1), json_object('b', 2))"), + Row("""[{"a":1},{"b":2}]""")) + } + + test("a nested JSON_QUERY value is spliced under KEEP QUOTES and quoted under OMIT QUOTES") { + // JSON_QUERY emits JSON text under the default KEEP QUOTES, so a lexically nested JSON_QUERY is + // spliced raw: the matched object is {"x":1}, not the quoted string "{\"x\":1}". + checkAnswer( + sql("""SELECT json_object('a' VALUE json_query('{"o":{"x":1}}', '$.o'))"""), + Row("""{"a":{"x":1}}""")) + // OMIT QUOTES returns the matched scalar string's decoded content (Ada, not "Ada") -- an + // ordinary string -- so it takes the quoted path (emitsImplicitJsonText is false), never the + // invalid splice {"a":Ada}. + checkAnswer( + sql("""SELECT json_object('a' VALUE json_query('{"n":"Ada"}', '$.n' OMIT QUOTES))"""), + Row("""{"a":"Ada"}""")) + } + + test("null key error") { + val e = intercept[SparkRuntimeException] { + sql("SELECT json_object(NULL VALUE 'value')").collect() + } + // Assert the structured error contract, not just the message text. + assert(e.getCondition == "JSON_OBJECT_NULL_KEY") + assert(e.getSqlState == "22004") + } + + test("a null key is validated before a null value is omitted under ABSENT ON NULL") { + // ABSENT ON NULL omits members with a null value, but the key is validated first, so a null key + // still raises JSON_OBJECT_NULL_KEY rather than being silently dropped along with the member. + val e = intercept[SparkRuntimeException] { + sql("SELECT json_object(NULL VALUE NULL ABSENT ON NULL)").collect() + } + assert(e.getCondition == "JSON_OBJECT_NULL_KEY") + assert(e.getSqlState == "22004") + } + + test("non-foldable key and value expressions") { + val df = Seq(("key1", "val1"), ("key2", "val2")).toDF("k", "v") + checkAnswer( + df.selectExpr("json_object(k VALUE v)"), + Seq(Row("""{"key1":"val1"}"""), Row("""{"key2":"val2"}"""))) + } + + test("non-foldable with NULL value and NULL ON NULL") { + val df = Seq(("k", null), ("key", "val")).toDF("k", "v") + checkAnswer( + df.selectExpr("json_object(k VALUE v)"), + Seq(Row("""{"k":null}"""), Row("""{"key":"val"}"""))) + } + + test("non-foldable with NULL value and ABSENT ON NULL") { + val df = Seq(("k", null), ("key", "val")).toDF("k", "v") + checkAnswer( + df.selectExpr("json_object(k VALUE v ABSENT ON NULL)"), + Seq(Row("{}"), Row("""{"key":"val"}"""))) + } + + test("multiple keys with ABSENT ON NULL") { + checkAnswer( + sql("""SELECT json_object('a' VALUE 1, 'b' VALUE NULL, 'c' VALUE 3 + ABSENT ON NULL)"""), + Row("""{"a":1,"c":3}""")) + } + + test("duplicate keys are emitted in source order") { + checkAnswer( + sql("SELECT json_object('k' VALUE 1, 'k' VALUE 2)"), + Row("""{"k":1,"k":2}""")) + } + + test("non-string key type is rejected at analysis, not at execution") { + val ex = intercept[AnalysisException] { + sql("SELECT json_object(1 VALUE 'x')") + } + assert(ex.getMessage.contains("UNEXPECTED_INPUT_TYPE")) + } + + test("non-string key type reports the actual key argument") { + val ex = intercept[AnalysisException] { + sql("SELECT json_object('ok' VALUE 1, 2 VALUE 'bad')") + } + checkError( + exception = ex, + condition = "DATATYPE_MISMATCH.UNEXPECTED_INPUT_TYPE", + sqlState = Some("42K09"), + parameters = Map( + "sqlExpr" -> "\"JSON_OBJECT(ok VALUE 1, 2 VALUE bad)\"", + "paramIndex" -> "third", + "requiredType" -> "\"STRING\"", + "inputSql" -> "\"2\"", + "inputType" -> "\"INT\""), + queryContext = Array(ExpectedContext("json_object('ok' VALUE 1, 2 VALUE 'bad')", 7, 46))) + } + + test("collated STRING RETURNING is accepted") { + // isValidReturningType must accept any StringType instance, not just the default collation. + checkAnswer( + sql("SELECT json_object('a' VALUE 1 RETURNING STRING COLLATE UTF8_LCASE)"), + Row("""{"a":1}""")) + } + + test("an invalid RETURNING type is reported under DATATYPE_MISMATCH") { + // The error is emitted as a DataTypeMismatch, so its condition must resolve under + // DATATYPE_MISMATCH -- not as a top-level INVALID_JSON_RETURNING_TYPE class. + val e = intercept[AnalysisException] { + sql("SELECT json_object('a' VALUE 1 RETURNING INT)").collect() + } + assert(e.getCondition == "DATATYPE_MISMATCH.INVALID_JSON_RETURNING_TYPE") + } + + test("a directly-constructed JsonObjectExpr with a CHAR/VARCHAR RETURNING is rejected") { + // The parser normalizes CHAR/VARCHAR RETURNING to STRING, but a raw CharType/VarcharType from + // direct Catalyst construction would advertise a length JSON_OBJECT does not enforce. + Seq(VarcharType(2), CharType(2)).foreach { returning => + val expr = JsonObjectExpr( + Seq((Literal("k"), Literal(1))), Seq(false), Seq(false), + JsonConstructorNullBehavior.Null, returning) + expr.checkInputDataTypes() match { + case DataTypeMismatch(errorSubClass, _) => + assert(errorSubClass == "INVALID_JSON_RETURNING_TYPE", s"for $returning") + case other => fail(s"expected DataTypeMismatch for $returning, got $other") + } + } + } + + test("value accepts an unparenthesized predicate expression") { + // valueExpr is parsed as a full `expression`, so ordinary predicates work without parentheses. + checkAnswer(sql("SELECT json_object('isnull' VALUE 1 IS NULL)"), Row("""{"isnull":false}""")) + checkAnswer(sql("SELECT json_object('gt' : 2 > 1)"), Row("""{"gt":true}""")) + } + + test("widening the value to expression does not change documented forms") { + // Design-doc examples where a value abuts the ON NULL / RETURNING keywords must still parse and + // evaluate identically after widening valueExpression -> expression. + checkAnswer(sql("SELECT json_object('id': 7, 'v': NULL)"), Row("""{"id":7,"v":null}""")) + checkAnswer( + sql("SELECT json_object('id': 7, 'v': NULL ABSENT ON NULL)"), Row("""{"id":7}""")) + checkAnswer( + sql("SELECT json_object('id', 7, 'v', NULL ABSENT ON NULL)"), Row("""{"id":7}""")) + checkAnswer( + sql("SELECT json_object('id' VALUE 7, 'name' VALUE 'Ada')"), + Row("""{"id":7,"name":"Ada"}""")) + } + + test("an unsupported value type is rejected at analysis") { + // A spatial value: JacksonUtils.verifyType accepts it (it is an AtomicType) but + // JacksonGenerator cannot serialize it, so JSON_OBJECT must reject it up front, not at runtime. + val bad = JsonObjectExpr( + Seq((Literal("k"), Literal.create(null, GeometryType(4326)))), + Seq(false), Seq(false), JsonConstructorNullBehavior.Null, StringType) + bad.checkInputDataTypes() match { + case DataTypeMismatch(sub, _) => assert(sub == "CANNOT_CONVERT_TO_JSON") + case other => fail(s"expected DataTypeMismatch, got $other") + } + // A spatial type nested as a MAP VALUE must still be rejected: the guard descends recursively + // into map values (unlike the top-level atomic case above), so this exercises that descent. + val badNested = JsonObjectExpr( + Seq((Literal("k"), Literal.create(null, MapType(StringType, GeometryType(4326))))), + Seq(false), Seq(false), JsonConstructorNullBehavior.Null, StringType) + badNested.checkInputDataTypes() match { + case DataTypeMismatch(sub, _) => assert(sub == "CANNOT_CONVERT_TO_JSON") + case other => fail(s"expected DataTypeMismatch for a nested spatial value, got $other") + } + // The guard also descends into STRUCT fields and ARRAY elements, so a spatial type buried in + // either -- the other two recursive branches -- must be rejected just like the map-value case. + val nestedBadTypes = Seq( + new StructType().add("g", GeometryType(4326)), + ArrayType(GeographyType(4326))) + nestedBadTypes.foreach { dt => + val expr = JsonObjectExpr( + Seq((Literal("k"), Literal.create(null, dt))), + Seq(false), Seq(false), JsonConstructorNullBehavior.Null, StringType) + expr.checkInputDataTypes() match { + case DataTypeMismatch(sub, _) => assert(sub == "CANNOT_CONVERT_TO_JSON", s"for $dt") + case other => fail(s"expected DataTypeMismatch for a spatial type in $dt, got $other") + } + } + // A spatial type appearing only as a MAP KEY is fine: JacksonGenerator writes map keys via + // toString, so the value-type guard must not over-reject it. + val ok = JsonObjectExpr( + Seq((Literal("k"), Literal.create(null, MapType(GeometryType(4326), IntegerType)))), + Seq(false), Seq(false), JsonConstructorNullBehavior.Null, StringType) + assert(ok.checkInputDataTypes().isSuccess) + } + + test("a raw (FORMAT JSON) value that is not a string is rejected") { + // A raw-spliced value must carry JSON text (string); a non-string raw value would fail with a + // ClassCastException at eval, so reject it at analysis with the FORMAT JSON input error. + val expr = JsonObjectExpr( + Seq((Literal("k"), Literal(1))), Seq(true), Seq(true), + JsonConstructorNullBehavior.Null, StringType) + expr.checkInputDataTypes() match { + case DataTypeMismatch(sub, _) => assert(sub == "INVALID_JSON_FORMAT_JSON_INPUT") + case other => fail(s"expected DataTypeMismatch, got $other") + } + } + + test("SQL renders an explicit collated RETURNING and an explicit ON NULL clause") { + val collated = JsonObjectExpr( + Seq((Literal("k"), Literal(1))), Seq(false), Seq(false), JsonConstructorNullBehavior.Null, + StringType("UTF8_LCASE")) + assert(collated.sql.contains("RETURNING STRING COLLATE UTF8_LCASE")) + // .sql always renders an explicit ON NULL clause so reparse stays on the direct grammar path. + val default = JsonObjectExpr( + Seq((Literal("k"), Literal(1))), Seq(false), Seq(false), + JsonConstructorNullBehavior.Null, StringType) + assert(default.sql == "JSON_OBJECT('k' VALUE 1 NULL ON NULL)") + } + + test("SQL renders a raw nested value spliced back to raw even after collation wrapping") { + val inner = JsonObjectExpr( + Seq((Literal("b"), Literal(1))), Seq(false), Seq(false), + JsonConstructorNullBehavior.Null, StringType) + // Simulate the default-collation rule wrapping the raw nested value in a Cast. rawJson stays + // frozen true; the Cast hides the bare constructor, so .sql renders an explicit FORMAT JSON so + // reparse splices it raw. + val wrapped = JsonObjectExpr( + Seq((Literal("a"), Cast(inner, StringType("UTF8_LCASE")))), Seq(true), Seq(false), + JsonConstructorNullBehavior.Null, StringType) + assert(wrapped.sql.contains("FORMAT JSON")) + checkAnswer(sql(s"SELECT ${wrapped.sql}"), Row("""{"a":{"b":1}}""")) + } + + test("emitted SQL reparses and evaluates with raw-vs-quoted semantics preserved") { + // The .sql renderings above are round-trip contracts: reparsing and evaluating them must + // reproduce the original raw-vs-quoted splicing. + // A raw nested value renders as a bare constructor and reparses back to raw splicing. + checkAnswer( + sql("SELECT JSON_OBJECT('a' VALUE JSON_OBJECT('b' VALUE 1))"), Row("""{"a":{"b":1}}""")) + // A quoted value that the optimizer inlined as an implicit-JSON expression is neutralized with + // CAST(... AS STRING); reparsing must keep it quoted rather than splicing it raw. + val inner = JsonObjectExpr( + Seq((Literal("b"), Literal(1))), Seq(false), Seq(false), + JsonConstructorNullBehavior.Null, StringType) + val quoted = JsonObjectExpr( + Seq((Literal("a"), inner)), Seq(false), Seq(false), + JsonConstructorNullBehavior.Null, StringType) + assert(quoted.sql == + "JSON_OBJECT('a' VALUE CAST(JSON_OBJECT('b' VALUE 1 NULL ON NULL) AS STRING) NULL ON NULL)") + checkAnswer(sql(s"SELECT ${quoted.sql}"), Row("""{"a":"{\"b\":1}"}""")) + } + + test("generated column names omit the synthetic NULL ON NULL ownership clause") { + // Canonical `.sql` appends a default NULL ON NULL so a clause-free call reparses to the builtin + // (see the shadowing test), but that marker is never reparsed for display. Unaliased column + // names must therefore not expose ownership text the user never wrote; a non-default + // ABSENT ON NULL the user did write is kept. + def colName(q: String): String = sql(q).schema.head.name + assert(colName("SELECT json_object()") == "JSON_OBJECT()") + assert(colName("SELECT json_object('id' VALUE 7)") == "JSON_OBJECT(id VALUE 7)") + assert(colName("SELECT json_object('id', 7)") == "JSON_OBJECT(id VALUE 7)") + assert(colName("SELECT json_object('a' VALUE json_object('b' VALUE 1))") == + "JSON_OBJECT(a VALUE JSON_OBJECT(b VALUE 1))") + assert(colName("SELECT json_object('v' VALUE NULL ABSENT ON NULL)") == + "JSON_OBJECT(v VALUE NULL ABSENT ON NULL)") + } + + test("ABSENT ON NULL round-trips through .sql keeping null members omitted") { + // The ABSENT renderer arm must emit `ABSENT ON NULL` (not the default clause and not nothing), + // so a canonical round trip keeps dropping null-valued members rather than retaining them. + val absent = JsonObjectExpr( + Seq((Literal("a"), Literal(1)), (Literal("b"), Literal.create(null, IntegerType))), + Seq(false, false), Seq(false, false), JsonConstructorNullBehavior.Absent, StringType) + assert(absent.sql == "JSON_OBJECT('a' VALUE 1, 'b' VALUE CAST(NULL AS INT) ABSENT ON NULL)") + checkAnswer(sql(s"SELECT ${absent.sql}"), Row("""{"a":1}""")) + } + + test("JSON_OBJECT is foldable only when it cannot throw") { + assert(JsonObjectExpr( + Seq((Literal("k"), Literal(1))), Seq(false), Seq(false), + JsonConstructorNullBehavior.Null, StringType).foldable) + // Folding a possibly-null key would surface the null-key error at optimization even for rows a + // filter/join drops. + assert(!JsonObjectExpr( + Seq((Literal.create(null, StringType), Literal(1))), Seq(false), Seq(false), + JsonConstructorNullBehavior.Null, StringType).foldable) + assert(!JsonObjectExpr( + Seq((Literal("k"), Literal("[1]"))), Seq(true), Seq(true), + JsonConstructorNullBehavior.Null, StringType).foldable) + } + + test("a constant JSON_OBJECT is accepted as a column DEFAULT") { + withTable("t") { + sql("CREATE TABLE t (id INT, c STRING DEFAULT json_object('a' VALUE 1)) USING parquet") + sql("INSERT INTO t (id) VALUES (1)") + checkAnswer(sql("SELECT c FROM t"), Row("""{"a":1}""")) + } + } + + test("a non-foldable JSON_OBJECT projection participates in whole-stage codegen") { + // Rendering through an Invoke of the object evaluator (not CodegenFallback) keeps the enclosing + // Project eligible for whole-stage codegen; a CodegenFallback expression would exclude it. A + // column-valued member keeps the constructor from being constant-folded away. + withSQLConf(SQLConf.ADAPTIVE_EXECUTION_ENABLED.key -> "false") { + val df = sql("SELECT json_object('id' VALUE id) FROM range(5)") + val fusedProjects = df.queryExecution.executedPlan.collect { + case w: WholeStageCodegenExec => w + }.flatMap(_.collect { case p: ProjectExec => p }) + assert(fusedProjects.nonEmpty, + s"expected the JSON_OBJECT Project to be fused into whole-stage codegen, but got:\n" + + df.queryExecution.executedPlan) + checkAnswer(df, (0 until 5).map(i => Row(s"""{"id":$i}"""))) + } + } + + test("a constant JSON_OBJECT is constant-folded") { + val df = sql("SELECT json_object('a' VALUE 1) AS r FROM range(2)") + assert(!df.queryExecution.optimizedPlan.exists( + _.expressions.exists(_.exists(_.isInstanceOf[JsonObjectExpr])))) + checkAnswer(df, Seq(Row("""{"a":1}"""), Row("""{"a":1}"""))) + } + + test("a null key raises JSON_OBJECT_NULL_KEY before the value is evaluated") { + // The key is checked before the value is evaluated, so a null key wins deterministically even + // when the value expression would itself throw. + // `raise_error(k)` references the column so it is neither foldable nor evaluated before the + // key null-check; if the value ran first the error would come from `raise_error`, not the key. + val e = intercept[SparkRuntimeException] { + sql("SELECT json_object(k VALUE raise_error(k)) " + + "FROM VALUES (CAST(NULL AS STRING)) t(k)").collect() + } + assert(e.getCondition == "JSON_OBJECT_NULL_KEY") + assert(e.getSqlState == "22004") + } + + test("a foldable literal key is rendered once and reused across rows") { + // JSON_OBJECT caches the rendered name of a foldable non-null key; the same key must still be Review Comment: Done — renamed the test and updated both comments to describe the evaluator's actual state (generator, writer, validated-text cache) instead of a key cache that no longer exists. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
