ganeshashree commented on code in PR #58034: URL: https://github.com/apache/spark/pull/58034#discussion_r4250579648
########## sql/core/src/test/scala/org/apache/spark/sql/JsonObjectSuite.scala: ########## @@ -0,0 +1,1211 @@ +/* + * 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.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 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 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 + // emitted for every row. + checkAnswer( + sql("SELECT json_object('id' VALUE a) FROM VALUES (1), (2) t(a)"), + Seq(Row("""{"id":1}"""), Row("""{"id":2}"""))) + // A foldable key that evaluates to null is not cached: it must still raise JSON_OBJECT_NULL_KEY + // per row rather than being silently skipped. + Seq("NULL", "CAST(NULL AS STRING)").foreach { k => + val e = intercept[SparkRuntimeException] { + sql(s"SELECT json_object($k VALUE 1)").collect() + } + assert(e.getCondition == "JSON_OBJECT_NULL_KEY", s"for key $k") + assert(e.getSqlState == "22004", s"for key $k") + } + } + + test("CHAR/VARCHAR RETURNING is normalized to STRING regardless of preserveCharVarcharTypeInfo") { + Seq("CHAR(2)", "VARCHAR(2)").foreach { returning => + Seq("true", "false").foreach { preserve => + withSQLConf(SQLConf.PRESERVE_CHAR_VARCHAR_TYPE_INFO.key -> preserve) { + assert( + sql(s"SELECT json_object('k' VALUE 1 RETURNING $returning)").schema.head.dataType + === StringType, + s"for RETURNING $returning, preserveCharVarcharTypeInfo=$preserve") + } + } + } + } + + test("object default collation applies only when RETURNING is not explicitly collated") { + withSQLConf(SQLConf.OBJECT_LEVEL_COLLATIONS_ENABLED.key -> "true") { + withTable("t") { + sql( + """CREATE TABLE t DEFAULT COLLATION UTF8_LCASE AS + |SELECT json_object('k' VALUE 1) AS a, + | json_object('k' VALUE 1 RETURNING STRING COLLATE UTF8_BINARY) AS b""".stripMargin) + val schema = spark.table("t").schema + // Omitted RETURNING (default STRING) follows the table's default collation. + assert(schema("a").dataType === StringType("UTF8_LCASE")) + // Explicit RETURNING ... COLLATE is the user's choice and must not be overwritten. + assert(schema("b").dataType === StringType("UTF8_BINARY")) + } + } + } + + test("default collation recurses into a nested JSON_OBJECT value") { + // The rule casts each DefaultStringProducingExpression, recursing through a nested constructor + // (the flat cases above only cover a top-level constructor). This CTAS runs the default + // analyzer (single-pass included). Confirm the schema collation and that raw splicing still + // produces well-formed nested JSON at runtime. + withSQLConf(SQLConf.OBJECT_LEVEL_COLLATIONS_ENABLED.key -> "true") { + withTable("t") { + sql( + """CREATE TABLE t DEFAULT COLLATION UTF8_LCASE AS + |SELECT json_object('a' VALUE json_object('b' VALUE 1)) AS a""".stripMargin) + assert(spark.table("t").schema("a").dataType === StringType("UTF8_LCASE")) + checkAnswer(spark.table("t"), Row("""{"a":{"b":1}}""")) + } + } + } + + test("view default collation preserves an explicit collated RETURNING") { + // Exercises the CREATE VIEW resolution path (in addition to the CTAS path above): the explicit + // RETURNING collation must survive the view's default collation. Runs under dual-run (the + // default) so the single-pass resolver's view re-resolution is exercised for parity. + withSQLConf(SQLConf.OBJECT_LEVEL_COLLATIONS_ENABLED.key -> "true") { + withView("v") { + sql( + """CREATE VIEW v DEFAULT COLLATION UTF8_LCASE AS + |SELECT json_object('k' VALUE 1) AS a, + | json_object('k' VALUE 1 RETURNING STRING COLLATE UTF8_BINARY) AS b""".stripMargin) + val schema = spark.table("v").schema + assert(schema("a").dataType === StringType("UTF8_LCASE")) + assert(schema("b").dataType === StringType("UTF8_BINARY")) + } + } + } + + test("reports nullable to keep NullPropagation from skipping the null-key check") { + // The value is never null, but a possibly-null key makes JSON_OBJECT throwable, so it must + // report nullable: were it non-nullable, `NullPropagation` would fold `IS [NOT] NULL` and + // `count(...)` away and skip the eval that must raise JSON_OBJECT_NULL_KEY (see below). + assert(JsonObjectExpr( + Seq((Literal.create(null, StringType), Literal(1))), Seq(false), Seq(false), + JsonConstructorNullBehavior.Null, StringType).nullable) + assert(sql("SELECT json_object(k VALUE 1) FROM VALUES ('a'), (NULL) t(k)") + .schema.head.nullable) + // A non-null key with a null value cannot throw, so the result is non-nullable. + assert(!JsonObjectExpr( + Seq((Literal("k"), Literal.create(null, IntegerType))), Seq(false), Seq(false), + JsonConstructorNullBehavior.Null, StringType).nullable) + assert(!sql("SELECT json_object('id' VALUE a) FROM VALUES (1), (2) t(a)") + .schema.head.nullable) + } + + test("IS NOT NULL and count still raise a null-key error") { + // NullPropagation must not rewrite these to a constant / `count(1)` and skip evaluation, which + // would swallow the null-key error on the null-key row. + Seq( + "SELECT json_object(k VALUE 1) IS NOT NULL FROM VALUES ('a'), (NULL) t(k)", + "SELECT count(json_object(k VALUE 1)) FROM VALUES ('a'), (NULL) t(k)").foreach { query => + val e = intercept[SparkRuntimeException](sql(query).collect()) + assert(e.getCondition == "JSON_OBJECT_NULL_KEY") + } + } + + test("is throwable only when a key may be null or a value needs validation") { + // A possibly-null key throws at runtime, so throwable must be true even when no child is + // itself throwable. + def obj(key: Expression, raw: Boolean = false): JsonObjectExpr = JsonObjectExpr( + Seq((key, Literal("[1]"))), Seq(raw), Seq(raw), JsonConstructorNullBehavior.Null, StringType) + assert(obj(Literal.create(null, StringType)).throwable) + assert(obj(Literal("k"), raw = true).throwable) + assert(!obj(Literal("k")).throwable) + } + + test("throwable keeps a null-key predicate above a filtering join") { + // The optimizer must not push a throwable predicate below the join (PushPredicateThroughJoin + // only pushes non-throwable conditions), so a null-key row the join eliminates is never + // evaluated and does not throw. Were the constructor not throwable, the predicate would push to + // the probe side and throw on the eliminated row. The predicate is a throwable comparison + // (`json_object(...) = literal`) that exercises the pushdown path. + withTempView("t", "u") { + Seq((1, "k1"), (2, null)).toDF("id", "k").createOrReplaceTempView("t") + Seq(1).toDF("id").createOrReplaceTempView("u") + // id=2 has a null key but does not join u, so it is dropped first (the predicate stays above + // the join, so the constructor is never evaluated on the eliminated row). + checkAnswer( + sql("""SELECT t.id FROM t JOIN u ON t.id = u.id + |WHERE json_object(t.k VALUE 1) = '{"k1":1}'""".stripMargin), + Row(1)) + // With the null-key row surviving the join, evaluation still throws. + Seq(2).toDF("id").createOrReplaceTempView("u") + val e = intercept[SparkRuntimeException] { + sql("""SELECT t.id FROM t JOIN u ON t.id = u.id + |WHERE json_object(t.k VALUE 1) = '{"k1":1}'""".stripMargin).collect() + } + assert(e.getCondition == "JSON_OBJECT_NULL_KEY") + } + } + + test("a non-throwing JSON_OBJECT predicate is pushed below a join") { + // Range-based views keep the filter from being folded into a local relation. + withTempView("t", "u") { + spark.range(3).selectExpr("id", "id AS x").createOrReplaceTempView("t") + spark.range(2).createOrReplaceTempView("u") + val df = sql("""SELECT t.id FROM t JOIN u ON t.id = u.id + |WHERE json_object('a' VALUE t.x) = '{"a":1}'""".stripMargin) + val join = df.queryExecution.optimizedPlan.collectFirst { case j: Join => j } + .getOrElse(fail("expected a Join in the optimized plan")) + assert(join.exists { + case f: Filter => f.condition.exists(_.isInstanceOf[JsonObjectExpr]) + case _ => false + }) + checkAnswer(df, Row(1)) + } + } + + test("raw splicing is decided from the source, not the optimized plan shape") { + // A directly-nested JSON_OBJECT carries implicit FORMAT JSON and is spliced raw. + checkAnswer( + sql("SELECT json_object('a' VALUE json_object('b' VALUE 1))"), + Row("""{"a":{"b":1}}""")) + // But a JSON_OBJECT result surfaced as a column is a plain STRING and must stay quoted -- even + // though CollapseProject may inline the inner JSON_OBJECT into the outer value position. The + // decision is frozen from the lexical argument at parse time, so the result stays + // {"a":"{\"b\":1}"}, never {"a":{"b":1}}. + val inlined = + sql("SELECT json_object('a' VALUE o) AS r FROM (SELECT json_object('b' VALUE 1) AS o) t") + checkAnswer(inlined, Row("""{"a":"{\"b\":1}"}""")) + // Referencing the alias twice blocks CollapseProject from inlining it; the result is identical, + // confirming independence from plan shape. + val notInlined = + sql("SELECT json_object('a' VALUE o) AS r, o FROM (SELECT json_object('b' VALUE 1) AS o) t") + checkAnswer(notInlined, Row("""{"a":"{\"b\":1}"}""", """{"b":1}""")) + } + + test("COLLATE on a nested JSON_OBJECT value still splices it raw") { + // COLLATE is a pass-through wrapper (collation metadata only), so a nested constructor under + // COLLATE is still spliced raw -- a no-op-looking annotation must not flip it to a quoted + // string, and the COLLATE form must agree with the bare form. + Seq("UTF8_BINARY", "UTF8_LCASE").foreach { collation => + checkAnswer( + sql(s"SELECT json_object('a' VALUE json_object('b' VALUE 1) COLLATE $collation)"), + Row("""{"a":{"b":1}}""")) + } + checkAnswer( + sql("SELECT json_object('a' VALUE json_object('b' VALUE 1) COLLATE UTF8_BINARY)"), + sql("SELECT json_object('a' VALUE json_object('b' VALUE 1))")) + } + + test("explicit CAST(... AS STRING) still cancels raw splicing") { + // In contrast to COLLATE, an explicit CAST(... AS STRING) is the documented way to quote a + // nested constructor rather than splice it raw. + checkAnswer( + sql("SELECT json_object('a' VALUE CAST(json_object('b' VALUE 1) AS STRING))"), + Row("""{"a":"{\"b\":1}"}""")) + } + + test("explicit FORMAT JSON splices a string value raw; a plain string is quoted") { + checkAnswer( + sql("""SELECT json_object('a' VALUE '{"b":1}' FORMAT JSON)"""), Row("""{"a":{"b":1}}""")) + checkAnswer( + sql("SELECT json_object('a' VALUE '[1,2]' FORMAT JSON)"), Row("""{"a":[1,2]}""")) + // FORMAT JSON applies to the colon form too, not just VALUE. + checkAnswer( + sql("""SELECT json_object('a' : '{"b":1}' FORMAT JSON)"""), Row("""{"a":{"b":1}}""")) + // Without FORMAT JSON the same string is quoted, even though its contents are valid JSON. + checkAnswer( + sql("""SELECT json_object('a' VALUE '{"b":1}')"""), Row("""{"a":"{\"b\":1}"}""")) + } + + test("per-member rawJson and validation flags stay tied to their own member") { + // Ordered object mixing a quoted string (rawJson=false), a trusted nested producer + // (rawJson=true), and an explicit FORMAT JSON string (rawJson=true, validated). Each member + // must render from its own flag, not a neighbour's. + checkAnswer( + sql("""SELECT json_object( + | 'a' VALUE '{"x":1}', + | 'b' VALUE json_object('c' VALUE 1), + | 'd' VALUE '[1,2]' FORMAT JSON)""".stripMargin), + Row("""{"a":"{\"x\":1}","b":{"c":1},"d":[1,2]}""")) + // A later malformed explicit value: validation must fire for that member (flattened arg 8), not + // be masked by the trusted/quoted earlier members. + val e = intercept[SparkRuntimeException] { + sql("""SELECT json_object( + | 'a' VALUE '{"x":1}', + | 'b' VALUE json_object('c' VALUE 1), + | 'd' VALUE '[1,2]' FORMAT JSON, + | 'e' VALUE '{bad' FORMAT JSON)""".stripMargin).collect() + } + assert(e.getCondition == "INVALID_JSON_FORMAT_JSON_VALUE") + assert(e.getMessageParameters.get("position") == "8") + } + + test("a malformed FORMAT JSON value raises an error at runtime") { + Seq("'{bad'", "'1,2'", "''").foreach { value => + val e = intercept[SparkRuntimeException] { + sql(s"SELECT json_object('a' VALUE $value FORMAT JSON)").collect() + } + assert(e.getCondition == "INVALID_JSON_FORMAT_JSON_VALUE", s"for $value") + } + // The reported position is the flattened value-argument position, not the member ordinal: the + // bad value below is the 4th argument (`a`, 1, `b`, `{bad`), not the 2nd member. + val runtimeErr = intercept[SparkRuntimeException] { + sql("SELECT json_object('a' VALUE 1, 'b' VALUE '{bad' FORMAT JSON)").collect() + } + assert(runtimeErr.getCondition == "INVALID_JSON_FORMAT_JSON_VALUE") + assert(runtimeErr.getMessageParameters.get("position") == "4") + } + + test("FORMAT JSON on a non-string value is rejected at analysis") { + val e = intercept[AnalysisException] { + sql("SELECT json_object('a' VALUE 123 FORMAT JSON)") + } + assert(e.getCondition == "DATATYPE_MISMATCH.INVALID_JSON_FORMAT_JSON_INPUT") + // The reported position is the flattened value-argument position, not the member ordinal: the + // bad value below is the 4th argument (`a`, 1, `b`, 123), not the 2nd member. + val laterMemberErr = intercept[AnalysisException] { + sql("SELECT json_object('a' VALUE 1, 'b' VALUE 123 FORMAT JSON)") + } + assert(laterMemberErr.getCondition == "DATATYPE_MISMATCH.INVALID_JSON_FORMAT_JSON_INPUT") + assert(laterMemberErr.getMessageParameters.get("position") == "4") + } + + test("a NULL FORMAT JSON value follows ON NULL like any other null") { + checkAnswer(sql("SELECT json_object('a' VALUE NULL FORMAT JSON)"), Row("""{"a":null}""")) + checkAnswer( + sql("SELECT json_object('a' VALUE NULL FORMAT JSON ABSENT ON NULL)"), Row("{}")) + } + + test("a per-row (non-foldable) FORMAT JSON value is validated for each row") { + // A column-valued FORMAT JSON is non-foldable, so it exercises the per-row validation branch + // (not the cached foldable branch). + checkAnswer( + Seq("""{"b":1}""").toDF("j").selectExpr("json_object('a' VALUE j FORMAT JSON)"), + Row("""{"a":{"b":1}}""")) + val e = intercept[SparkRuntimeException] { + Seq("bad").toDF("j").selectExpr("json_object('a' VALUE j FORMAT JSON)").collect() + } + assert(e.getCondition == "INVALID_JSON_FORMAT_JSON_VALUE") + } + + test("a constant-folded raw nested value round-trips through .sql as raw") { + // ConstantFolding rewrites the nested json_array(1) to the literal '[1]', but rawJson stays + // frozen true. .sql must emit FORMAT JSON so reparse splices it raw rather than quoting it. The + // column-valued member keeps the outer constructor from folding. + val from = "FROM VALUES (1) t(x)" + val jsonObj = sql(s"SELECT json_object('a' VALUE json_array(1), 'b' VALUE x) AS r $from") + .queryExecution.optimizedPlan.expressions + .flatMap(_.collect { case j: JsonObjectExpr => j }).head + assert(jsonObj.children(1).isInstanceOf[Literal], + "the nested constructor should have been constant-folded to a literal") + assert(jsonObj.sql.contains("FORMAT JSON")) + checkAnswer(sql(s"SELECT ${jsonObj.sql} AS r $from"), Row("""{"a":[1],"b":1}""")) + } + + test("canonical .sql reparses to the built-in even under a shadowing routine") { + // .sql renders an explicit ON NULL clause, forcing the direct grammar path, so a resolved + // built-in round-trips to the built-in even when a same-named routine is on the path. + val objSql = JsonObjectExpr( + Seq((Literal("id"), Literal(7))), Seq(false), Seq(false), + JsonConstructorNullBehavior.Null, StringType).sql + assert(objSql.contains("NULL ON NULL")) + withSQLConf( + SQLConf.PATH_ENABLED.key -> "true", + SQLConf.SESSION_FUNCTION_RESOLUTION_ORDER.key -> "second") { + try { + sql("CREATE TEMPORARY FUNCTION json_object(a STRING, b INT) RETURNS STRING " + + "RETURN 'shadowed'") + sql("SET PATH = system.session, system.builtin") + // The clause-free call is shadowed, but the canonical .sql (with ON NULL) is not. + checkAnswer(sql("SELECT json_object('id', 7)"), Row("shadowed")) + checkAnswer(sql(s"SELECT $objSql"), Row("""{"id":7}""")) + } finally { + sql("SET PATH = DEFAULT_PATH") + sql("DROP TEMPORARY FUNCTION IF EXISTS json_object") + } + } + } + + test("SQL renders a COLLATE-wrapped raw nested value as its bare constructor") { + // A raw value behind a pass-through Collate renders as the bare inner constructor so reparse + // re-derives raw splicing (the collation does not affect the spliced-raw bytes). + val inner = JsonObjectExpr( + Seq((Literal("b"), Literal(1))), Seq(false), Seq(false), + JsonConstructorNullBehavior.Null, StringType) + val wrapped = JsonObjectExpr( + Seq((Literal("a"), Collate(inner, Literal("UTF8_BINARY")))), Seq(true), Seq(false), + JsonConstructorNullBehavior.Null, StringType) + assert(wrapped.sql == + "JSON_OBJECT('a' VALUE JSON_OBJECT('b' VALUE 1 NULL ON NULL) NULL ON NULL)") + checkAnswer(sql(s"SELECT ${wrapped.sql}"), Row("""{"a":{"b":1}}""")) + } + + test("plain call goes through routine resolution and can be shadowed via SET PATH") { + // `withUserDefinedFunction` is unusable here: its cleanup asserts the name no longer resolves, + // but `json_object` is now a registered built-in, so drop the temporary routine explicitly. + withSQLConf( + SQLConf.PATH_ENABLED.key -> "true", + SQLConf.SESSION_FUNCTION_RESOLUTION_ORDER.key -> "second") { + try { + sql("CREATE TEMPORARY FUNCTION json_object(a STRING, b INT) RETURNS STRING " + + "RETURN 'shadowed'") + sql("SET PATH = system.session, system.builtin") + // A plain call is an ordinary function call, so the temporary routine (ahead of + // system.builtin on the path) shadows the built-in constructor. + checkAnswer(sql("SELECT json_object('id', 7)"), Row("shadowed")) + checkAnswer(sql("SELECT json_object('id' VALUE 7)"), Row("shadowed")) Review Comment: Done. ########## sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/jsonExpressions.scala: ########## @@ -2272,3 +2331,424 @@ case class JsonTypeof(child: Expression) override protected def withNewChildInternal(newChild: Expression): JsonTypeof = copy(child = newChild) } + +/** + * The SQL:2016 `JSON_OBJECT` constructor function (feature T811): constructs a JSON object from + * key-value pairs, written `key VALUE value`, `KEY key VALUE value`, `key : value`, or the + * MySQL-style `key, value`. + * + * Keys must be non-null strings; a null key is an error. The `ON NULL` clause controls whether + * null-valued pairs are included (NULL ON NULL, the standard default) or omitted (ABSENT ON NULL). + * + * A value is spliced in raw (unquoted) when it carries `FORMAT JSON` (tracked per member by + * `rawJson`): the explicit clause (e.g. `JSON_OBJECT('a' VALUE '{"b":1}' FORMAT JSON)`), or the + * implicit form carried by a nested JSON constructor / `JSON_QUERY` under KEEP QUOTES (see + * `rawJsonValue`). Explicit `FORMAT JSON` on a non-constructor is arbitrary user text, so + * `needsValidation` marks it for JSON validation at eval; a nested constructor is trusted. + * + * Examples: + * JSON_OBJECT('id' VALUE 7, 'name' VALUE 'Ada') -> '{"id":7,"name":"Ada"}' + * JSON_OBJECT('id': 7, 'v': NULL) -> '{"id":7,"v":null}' + * JSON_OBJECT('id': 7, 'v': NULL ABSENT ON NULL) -> '{"id":7}' + * JSON_OBJECT() -> '{}' + * + * A clause-free call that is not itself lexically nested in an enclosing constructor's *value* + * position routes through function resolution and is rebuilt by `JsonObjectExpressionBuilder`, so a + * same-named routine can shadow the built-in `json_object`: always for the comma form (a nested + * producer value routes too, carrying its raw-splice eligibility via `JsonImplicitFormatCarrier`), + * but for the `VALUE` / `:` forms only when no member carries a raw value (`rawJson.contains(true)` + * -- an explicit `FORMAT JSON` or a nested producer in *value* position -- keeps it direct). + * Otherwise it is built directly from the grammar (clause-bearing, or value-nested); a nested + * *key*-position call stays routable (see `AstBuilder.visitJsonObject`). + * The user-facing reference lives in `docs/sql-ref-syntax-qry-select-json-object.md`. + */ +case class JsonObjectExpr( + members: Seq[(Expression, Expression)], + rawJson: Seq[Boolean], + needsValidation: Seq[Boolean], + nullBehavior: JsonConstructorNullBehavior = JsonConstructorNullBehavior.Null, + returning: DataType = StringType, + timeZoneId: Option[String] = None) + extends Expression + with TimeZoneAwareExpression + with CodegenFallback + with QueryErrorsBase + // Default RETURNING is a plain STRING, so `DefaultStringProducingExpression` lets + // `ApplyDefaultCollation` cast the result to a non-default collation; the `dataType` override + // below stays authoritative only for an explicitly collated RETURNING. + with DefaultStringProducingExpression + with ImplicitlyFormattedAsJson + with RoutedSqlJsonExpression { + + // `rawJson(i)` splices member `i`'s value verbatim (already-JSON text) rather than quoting it; + // `needsValidation(i)` marks that raw text as arbitrary user input to JSON-validate at eval (an + // explicit `FORMAT JSON` on a non-constructor). Both are frozen at parse time (see + // `AstBuilder.visitJsonObject`). A validated value is always a spliced one. + require(members.length == rawJson.length && members.length == needsValidation.length, + "JsonObjectExpr requires one rawJson and one needsValidation flag per member") + require(needsValidation.lazyZip(rawJson).forall((nv, raw) => !nv || raw), + "JsonObjectExpr needsValidation implies rawJson") + + @transient private lazy val memberArray: Array[(Expression, Expression)] = + members.toArray + + @transient private lazy val rawJsonArray: Array[Boolean] = + rawJson.toArray + + @transient private lazy val needsValidationArray: Array[Boolean] = + needsValidation.toArray + + // Throwable only when eval can throw: a key that may be null (JSON_OBJECT_NULL_KEY), an explicit + // `FORMAT JSON` value to validate, or a throwable child. This keeps such shapes from being pushed + // below a filtering join. + override lazy val throwable: Boolean = + children.exists(_.throwable) || needsValidation.contains(true) || + members.exists { case (k, _) => !k.resolved || k.nullable } + + // The value is never null, but report nullable as `throwable` (like JsonArray): `NullPropagation` + // keys off `nullable`, so a non-nullable constructor would let it fold `IS [NOT] NULL` to a + // constant or `count(...)` to `count(1)`, skipping the eval that must raise JSON_OBJECT_NULL_KEY. + override def nullable: Boolean = throwable + + // Only non-throwing shapes fold, so `ConstantFolding` cannot raise an error at optimization for + // rows a filter would later drop. + override def foldable: Boolean = children.forall(_.foldable) && !throwable + + override def dataType: DataType = returning + + // Flattened `[k0, v0, k1, v1, ...]` view, cached since `children` is walked repeatedly. + @transient private lazy val childrenSeq: Seq[Expression] = + memberArray.flatMap { case (k, v) => Seq(k, v) }.toImmutableArraySeq + + override def children: Seq[Expression] = childrenSeq + + override def withTimeZone(timeZoneId: String): TimeZoneAwareExpression = + copy(timeZoneId = Option(timeZoneId)) + + override def checkInputDataTypes(): TypeCheckResult = { + val inputCheck = super.checkInputDataTypes() + if (inputCheck.isFailure) { + inputCheck + } else if (!JsonObjectExpr.isValidReturningType(returning)) { + DataTypeMismatch( + errorSubClass = "INVALID_JSON_RETURNING_TYPE", + messageParameters = Map( + "functionName" -> toSQLId(prettyName), + "returningType" -> toSQLType(returning))) + } else { + // Keys must be character strings (they become JSON object member names). Reject any other + // concrete type up front so a non-string key cannot reach the StringType Jackson writer and + // fail with an internal ClassCastException. NullType is allowed through: a null key value is + // reported at runtime as JSON_OBJECT_NULL_KEY rather than a type error. + val keyCheck = memberArray.iterator.map(_._1).zipWithIndex.collectFirst { + case (k, keyIndex) if !(k.dataType.isInstanceOf[StringType] || k.dataType == NullType) => + DataTypeMismatch( + errorSubClass = "UNEXPECTED_INPUT_TYPE", + messageParameters = Map( + "paramIndex" -> ordinalNumber(keyIndex * 2), + "requiredType" -> toSQLType(StringType), + "inputSql" -> toSQLExpr(k), + "inputType" -> toSQLType(k.dataType))) + }.getOrElse(TypeCheckResult.TypeCheckSuccess) + if (keyCheck.isFailure) { + keyCheck + } else { + // Every value must be serializable to JSON. This mirrors `to_json`'s analysis-time + // `JacksonUtils.verifyType` check, plus a guard for the spatial atomics that check accepts + // but `JacksonGenerator` cannot write (rejecting a runtime failure at analysis instead). + memberArray.iterator.map(_._2).zipWithIndex + .foldLeft(TypeCheckResult.TypeCheckSuccess: TypeCheckResult) { + case (acc, _) if acc.isFailure => acc + case (_, (v, valueIndex)) if rawJsonArray(valueIndex) && + !(v.dataType.isInstanceOf[StringType] || v.dataType == NullType) => + // A raw-spliced value must carry JSON text (string); NullType is allowed (eval drops + // or emits it before splicing). Reject any other type up front, matching JSON_ARRAY's + // FORMAT JSON check, rather than failing with a ClassCastException at eval. + DataTypeMismatch( + errorSubClass = "INVALID_JSON_FORMAT_JSON_INPUT", + messageParameters = Map( + "functionName" -> toSQLId(prettyName), + // Flattened 1-based arg position, not the member ordinal (value of member i is + // arg 2i+2), matching the key check's `ordinalNumber(keyIndex*2)`. + "position" -> (valueIndex * 2 + 2).toString, + "inputType" -> toSQLType(v.dataType))) + case (_, (v, _)) => + val elemCheck = JacksonUtils.verifyType(prettyName, v.dataType) + if (elemCheck.isFailure) { + elemCheck + } else if (JsonObjectExpr.containsUnsupportedJsonType(v.dataType)) { + DataTypeMismatch( + errorSubClass = "CANNOT_CONVERT_TO_JSON", + messageParameters = Map( + "name" -> toSQLId(prettyName), + "type" -> toSQLType(v.dataType))) + } else { + TypeCheckResult.TypeCheckSuccess + } + } + } + } + } + + override def stateful: Boolean = true + + @transient private lazy val resolvedZoneId: String = + timeZoneId.getOrElse(SQLConf.get.sessionLocalTimeZone) + + // JacksonGenerator (shared with `to_json`, so rendering matches) can only serialize a container, + // not a bare scalar, so each key/value is wrapped in a one-element array whose brackets are + // stripped in renderKey/appendRenderedValue below. + @transient private lazy val keyEvaluator: StructsToJsonEvaluator = + StructsToJsonEvaluator(Map.empty, ArrayType(StringType), Some(resolvedZoneId)) + + @transient private lazy val valueEvaluators: Array[StructsToJsonEvaluator] = + new Array[StructsToJsonEvaluator](memberArray.length) + + @transient private lazy val singleElem: Array[Any] = new Array[Any](1) + + @transient private lazy val singleElemData: GenericArrayData = + new GenericArrayData(singleElem) + + @transient private lazy val castInput: GenericInternalRow = + new GenericInternalRow(1) + + @transient private lazy val returningCast: Expression = + Cast(BoundReference(0, StringType, nullable = true), returning, timeZoneId, + EvalMode.ANSI) + + // `evaluateString` returns the Java String directly, skipping `evaluate`'s UTF8String round trip. + // The key is cached or appended whole, so it still materializes the interior substring. + private def renderKey(key: Any): String = { + singleElem(0) = key + val arrJson = keyEvaluator.evaluateString(singleElemData) + arrJson.substring(1, arrJson.length - 1) + } + + // Append the interior of the "[<frag>]" wrapper straight into the builder (no per-value + // substring), mirroring `JsonArray.appendRenderedElement`. + private def appendRenderedValue(sb: java.lang.StringBuilder, idx: Int, value: Any): Unit = { + singleElem(0) = value + var evaluator = valueEvaluators(idx) + if (evaluator == null) { + evaluator = StructsToJsonEvaluator( + Map.empty, ArrayType(memberArray(idx)._2.dataType), Some(resolvedZoneId)) + valueEvaluators(idx) = evaluator + } + val arrJson = evaluator.evaluateString(singleElemData) + sb.append(arrJson, 1, arrJson.length - 1) + } + + @transient private lazy val formatJsonFactory: JsonFactory = new JsonFactory() + + @transient private lazy val cachedValidatedTexts: Array[String] = + new Array[String](memberArray.length) + + // A `FORMAT JSON` value is spliced verbatim, so it must be exactly one well-formed JSON value + // (a string like `1,2` or `{bad` would corrupt the object). Parse one value; reject anything + // trailing. Mirrors `JsonArray.validateJsonText`. + private def validateJsonText(idx: Int, text: String): Unit = { + val valid = + try { + Utils.tryWithResource(formatJsonFactory.createParser(text)) { parser => + if (parser.nextToken() == null) { + false + } else { + parser.skipChildren() + parser.nextToken() == null + } + } + } catch { + case _: JsonProcessingException => false + } + if (!valid) { + // Flattened 1-based value-arg position (member `idx` is arg `2*idx+2`), matching the + // analysis-time INVALID_JSON_FORMAT_JSON_INPUT check. + throw QueryExecutionErrors.invalidJsonFormatJsonValueError(prettyName, idx * 2 + 2, text) + } + } + + // Raw text to splice for member `idx`: a trusted nested constructor splices as-is; an explicit + // `FORMAT JSON` string is validated once (cached) when foldable, else per row. + private def rawJsonText(idx: Int, value: Any): String = { + if (!needsValidationArray(idx)) { + value.asInstanceOf[UTF8String].toString + } else if (memberArray(idx)._2.foldable) { Review Comment: Done. -- 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]
