[ https://issues.apache.org/jira/browse/FLINK-10019?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16578793#comment-16578793 ]
Rong Rong edited comment on FLINK-10019 at 8/14/18 5:39 PM: ------------------------------------------------------------ Dug a little deeper in Calcite and found that in: https://github.com/apache/calcite/blob/master/core/src/main/java/org/apache/calcite/sql/type/InferTypes.java#L68 Seems like if the return type is inferred as a record, or *{{.isStruct() == true}}*, "it must have the same number of fields as the number of operands." which clearly is not the case here since the following expression: *{{AS(func(a), "myRow")}}* only passes over the *{{func(a)}}* for type inference, but not the alias *{{"myRow"}}* was (Author: walterddr): Dug a little deeper in Calcite and found that in: https://github.com/apache/calcite/blob/master/core/src/main/java/org/apache/calcite/sql/type/InferTypes.java#L68 Seems like if the return type is inferred as a record, or {{isStruct()}}, "it must have the same number of fields as the number of operands." which clearly is not the case here since the following expression: **{{AS(func(a), "myRow")}}** only passes over the **{{func(a)}}** for type inference, but not the alias **{{"myRow"}}** > Fix Composite getResultType of UDF cannot be chained with other operators > ------------------------------------------------------------------------- > > Key: FLINK-10019 > URL: https://issues.apache.org/jira/browse/FLINK-10019 > Project: Flink > Issue Type: Bug > Reporter: Rong Rong > Assignee: Rong Rong > Priority: Major > > If explicitly return a CompositeType in {{udf.getResultType}}, will result in > some failures in chained operators. > For example: consider a simple UDF, > {code:scala} > object Func extends ScalarFunction { > def eval(row: Row): Row = { > row > } > override def getParameterTypes(signature: Array[Class[_]]): > Array[TypeInformation[_]] = > Array(Types.ROW(Types.INT)) > override def getResultType(signature: Array[Class[_]]): TypeInformation[_] = > Types.ROW(Types.INT) > } > {code} > This should work perfectly since it's just a simple pass through, however > {code:scala} > @Test > def testRowType(): Unit = { > val data = List( > Row.of(Row.of(12.asInstanceOf[Integer]), "1") > ) > val env = StreamExecutionEnvironment.getExecutionEnvironment > val stream = env.fromCollection(data)(Types.ROW(Types.ROW(Types.INT), > Types.STRING)) > val tEnv = TableEnvironment.getTableEnvironment(env) > val table = stream.toTable(tEnv, 'a, 'b) > tEnv.registerFunction("func", Func) > tEnv.registerTable("t", table) > // This works perfectly > val result1 = tEnv.sqlQuery("SELECT func(a) FROM t").toAppendStream[Row] > result1.addSink(new StreamITCase.StringSink[Row]) > // This throws exception > val result2 = tEnv.sqlQuery("SELECT func(a) as myRow FROM > t").toAppendStream[Row] > result2.addSink(new StreamITCase.StringSink[Row]) > env.execute() > } > {code} > Exception code: > {code:java} > java.lang.IndexOutOfBoundsException: index (1) must be less than size (1) > at > com.google.common.base.Preconditions.checkElementIndex(Preconditions.java:310) > at > com.google.common.base.Preconditions.checkElementIndex(Preconditions.java:293) > at > com.google.common.collect.SingletonImmutableList.get(SingletonImmutableList.java:41) > at > org.apache.calcite.sql.type.InferTypes$2.inferOperandTypes(InferTypes.java:83) > at > org.apache.calcite.sql.validate.SqlValidatorImpl.inferUnknownTypes(SqlValidatorImpl.java:1777) > at > org.apache.calcite.sql.validate.SqlValidatorImpl.expandSelectItem(SqlValidatorImpl.java:459) > at > org.apache.calcite.sql.validate.SqlValidatorImpl.expandStar(SqlValidatorImpl.java:349) > ... > {code} > This is due to the fact that Calcite inferOperandTypes does not expect to > infer a struct RelDataType. -- This message was sent by Atlassian JIRA (v7.6.3#76005)