szehon-ho commented on code in PR #55970: URL: https://github.com/apache/spark/pull/55970#discussion_r3269819036
########## sql/pipelines/src/main/scala/org/apache/spark/sql/pipelines/autocdc/Scd1BatchProcessor.scala: ########## @@ -0,0 +1,161 @@ +/* + * 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.pipelines.autocdc + +import org.apache.spark.SparkException +import org.apache.spark.sql.{functions => F, AnalysisException} +import org.apache.spark.sql.Column +import org.apache.spark.sql.catalyst.util.QuotingUtils +import org.apache.spark.sql.classic.DataFrame +import org.apache.spark.sql.types.{DataType, StructField, StructType} +import org.apache.spark.util.ArrayImplicits._ + +/** + * Per-microbatch processor for SCD Type 1 AutoCDC flows, complying to the specified [[changeArgs]] + * configuration. + * + * @param changeArgs The CDC flow configuration. + * @param resolvedSequencingType The post-analysis [[DataType]] of the sequencing column, derived + * from the flow's resolved DataFrame at flow setup time. + */ +case class Scd1BatchProcessor( + changeArgs: ChangeArgs, + resolvedSequencingType: DataType) { + + /** + * Deduplicate the incoming CDC microbatch by key, keeping the most recent event per key + * as ordered by [[ChangeArgs.sequencing]]. + * + * For SCD1 we only care about the most recent (by sequence value) event per key. When + * multiple events share the same key and the same sequence value, the row selected is + * non-deterministic and undefined. + * + * The schema of the returned dataframe matches the schema of the microbatch exactly. + */ + def deduplicateMicrobatch(microbatchDf: DataFrame): DataFrame = { + // The `max_by` API can only return a single column, so pack/unpack the entire row into a + // temporary column before and after the `max_by` operation. + val winningRowCol = OutOfOrderCdcMergeUtils.tempColName("__winning_row") + + val allMicrobatchColumns = + microbatchDf.columns + .map(colName => F.col(QuotingUtils.quoteIdentifier(colName))) + .toImmutableArraySeq + + microbatchDf + .groupBy(changeArgs.keys.map(k => F.col(k.quoted)): _*) + .agg( + F.max_by(F.struct(allMicrobatchColumns: _*), changeArgs.sequencing) + .as(winningRowCol) + ) + .select(F.col(s"$winningRowCol.*")) + } + + /** + * Project the CDC metadata column onto the microbatch. + * + * The returned dataframe has all of the columns in the input microbatch + the CDC metadata + * column. + */ + def extendMicrobatchRowsWithCdcMetadata(microbatchDf: DataFrame): DataFrame = { + // Proactively validate the reserved CDC metadata column does not exist in the microbatch. + validateCdcMetadataColumnNotPresent(microbatchDf) + + val rowDeleteSequence: Column = changeArgs.deleteCondition match { + case Some(deleteCondition) => + F.when(deleteCondition, changeArgs.sequencing).otherwise(F.lit(null)) + case None => + F.lit(null) + } + + val rowUpsertSequence: Column = + // A row that is not a delete must be an upsert, these are mutually exclusive and a complete + // set of CDC event types. + F.when(rowDeleteSequence.isNull, changeArgs.sequencing).otherwise(F.lit(null)) Review Comment: When `deleteCondition` evaluates to **NULL** (unknown), `rowDeleteSequence` is NULL and `rowDeleteSequence.isNull` is true, so the row gets `upsertSequence` (classified as upsert). That's probably what you want ("only explicit deletes count"), but worth a one-line scaladoc note or a small test so it's intentional rather than accidental. ########## sql/pipelines/src/main/scala/org/apache/spark/sql/pipelines/autocdc/Scd1BatchProcessor.scala: ########## @@ -0,0 +1,161 @@ +/* + * 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.pipelines.autocdc + +import org.apache.spark.SparkException +import org.apache.spark.sql.{functions => F, AnalysisException} +import org.apache.spark.sql.Column +import org.apache.spark.sql.catalyst.util.QuotingUtils +import org.apache.spark.sql.classic.DataFrame +import org.apache.spark.sql.types.{DataType, StructField, StructType} +import org.apache.spark.util.ArrayImplicits._ + +/** + * Per-microbatch processor for SCD Type 1 AutoCDC flows, complying to the specified [[changeArgs]] + * configuration. + * + * @param changeArgs The CDC flow configuration. + * @param resolvedSequencingType The post-analysis [[DataType]] of the sequencing column, derived + * from the flow's resolved DataFrame at flow setup time. + */ +case class Scd1BatchProcessor( + changeArgs: ChangeArgs, + resolvedSequencingType: DataType) { + + /** + * Deduplicate the incoming CDC microbatch by key, keeping the most recent event per key + * as ordered by [[ChangeArgs.sequencing]]. + * + * For SCD1 we only care about the most recent (by sequence value) event per key. When + * multiple events share the same key and the same sequence value, the row selected is + * non-deterministic and undefined. + * + * The schema of the returned dataframe matches the schema of the microbatch exactly. + */ + def deduplicateMicrobatch(microbatchDf: DataFrame): DataFrame = { Review Comment: Optional: no test yet for microbatches with nested **column types** (e.g. a top-level `struct` payload). The `struct` + `max_by` + `.*` unpack pattern should carry nested values from the winning row atomically; a small test would lock that in if sources commonly have struct columns. (Nested **paths** as keys remain out of scope per `UnqualifiedColumnName`.) ########## sql/pipelines/src/test/scala/org/apache/spark/sql/pipelines/autocdc/Scd1BatchProcessorSuite.scala: ########## @@ -0,0 +1,443 @@ +/* + * 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.pipelines.autocdc + +import org.apache.spark.SparkFunSuite +import org.apache.spark.sql.{functions => F, AnalysisException, Row} +import org.apache.spark.sql.classic.DataFrame +import org.apache.spark.sql.internal.SQLConf +import org.apache.spark.sql.test.SharedSparkSession +import org.apache.spark.sql.types._ + +class Scd1BatchProcessorSuite extends SparkFunSuite with SharedSparkSession { + + /** Build a microbatch [[DataFrame]] from explicit rows and an explicit schema. */ + private def microbatchOf(schema: StructType)(rows: Row*): DataFrame = + spark.createDataFrame(spark.sparkContext.parallelize(rows), schema) + + /** + * Returns the `(name, dataType)` pairs of `schema`'s fields. Used to compare two schemas for + * structural equivalence while deliberately ignoring nullability and metadata, which can shift + * benignly when columns are unpacked from a struct. + */ + private def columnNamesAndDataTypes(schema: StructType): Seq[(String, DataType)] = + schema.fields.map(f => (f.name, f.dataType)).toSeq + + test("deduplicateMicrobatch keeps only the row with the largest sequence value per key") { + val schema = new StructType() + .add("id", IntegerType) + .add("seq", LongType) + .add("value", StringType) + + val batch = microbatchOf(schema)( + Row(1, 10L, "first"), + Row(1, 30L, "winner"), + Row(1, 20L, "middle") + ) + + val processor = Scd1BatchProcessor( + changeArgs = ChangeArgs( + keys = Seq(UnqualifiedColumnName("id")), + sequencing = F.col("seq"), + storedAsScdType = ScdType.Type1 + ), + resolvedSequencingType = LongType + ) + + checkAnswer( + df = processor.deduplicateMicrobatch(batch), + expectedAnswer = Row(1, 30L, "winner") + ) + } + + test("deduplicateMicrobatch processes multiple keys independently") { + val schema = new StructType() + .add("id", IntegerType) + .add("seq", LongType) + .add("value", StringType) + + val batch = microbatchOf(schema)( + Row(1, 10L, "a1"), + Row(2, 50L, "b1-winner"), + Row(1, 20L, "a2-winner"), + Row(2, 40L, "b2-loser"), + Row(3, 1L, "c1-only") + ) + + val processor = Scd1BatchProcessor( + changeArgs = ChangeArgs( + keys = Seq(UnqualifiedColumnName("id")), + sequencing = F.col("seq"), + storedAsScdType = ScdType.Type1 + ), + resolvedSequencingType = LongType + ) + + checkAnswer( + df = processor.deduplicateMicrobatch(batch), + expectedAnswer = Seq( + Row(1, 20L, "a2-winner"), + Row(2, 50L, "b1-winner"), + Row(3, 1L, "c1-only") + ) + ) + } + + test("deduplicateMicrobatch carries non-key, non-sequence columns from the winning row") { + val schema = new StructType() + .add("id", IntegerType) + .add("seq", LongType) + .add("name", StringType) + .add("amount", IntegerType) + + val batch = microbatchOf(schema)( + Row(1, 10L, "old-name", 100), + Row(1, 20L, "winning-name", 200) + ) + + val processor = Scd1BatchProcessor( + changeArgs = ChangeArgs( + keys = Seq(UnqualifiedColumnName("id")), + sequencing = F.col("seq"), + storedAsScdType = ScdType.Type1 + ), + resolvedSequencingType = LongType + ) + + // All non-key columns must come from the row with the largest sequence value, never + // a mix of values from multiple rows. + checkAnswer( + df = processor.deduplicateMicrobatch(batch), + expectedAnswer = Row(1, 20L, "winning-name", 200) + ) + } + + test("deduplicateMicrobatch supports composite (multi-column) keys") { + val schema = new StructType() + .add("region", StringType) + .add("customer_id", IntegerType) + .add("seq", LongType) + .add("value", StringType) + + val batch = microbatchOf(schema)( + Row("US", 1, 10L, "us1-old"), + Row("US", 1, 20L, "us1-new"), + // Same customer_id as above but different region: independent group. + Row("EU", 1, 5L, "eu1-only"), + // Same region as above but different customer_id: independent group. + Row("US", 2, 99L, "us2-only") + ) + + val processor = Scd1BatchProcessor( + changeArgs = ChangeArgs( + keys = Seq(UnqualifiedColumnName("region"), UnqualifiedColumnName("customer_id")), + sequencing = F.col("seq"), + storedAsScdType = ScdType.Type1 + ), + resolvedSequencingType = LongType + ) + + checkAnswer( + df = processor.deduplicateMicrobatch(batch), + expectedAnswer = Seq( + Row("US", 1, 20L, "us1-new"), + Row("EU", 1, 5L, "eu1-only"), + Row("US", 2, 99L, "us2-only") + ) + ) + } + + test("deduplicateMicrobatch supports literal-dot column names") { + val schema = new StructType() + .add("user.id", IntegerType) + .add("seq", LongType) + .add("event.value", StringType) + + val batch = microbatchOf(schema)( + Row(1, 10L, "old"), + Row(1, 20L, "new") + ) + + val processor = Scd1BatchProcessor( + changeArgs = ChangeArgs( + keys = Seq(UnqualifiedColumnName("`user.id`")), + sequencing = F.col("seq"), + storedAsScdType = ScdType.Type1 + ), + resolvedSequencingType = LongType + ) + + checkAnswer( + df = processor.deduplicateMicrobatch(batch), + expectedAnswer = Row(1, 20L, "new") + ) + } + + test("deduplicateMicrobatch preserves the input column names, types, and ordering") { + val schema = new StructType() + .add("a", StringType) + .add("id", IntegerType) + .add("z", DoubleType) + .add("seq", LongType) + .add("flag", BooleanType) + + val batch = microbatchOf(schema)( + Row("a1", 1, 1.5, 10L, true), + Row("a2", 1, 2.5, 20L, false) + ) + + val processor = Scd1BatchProcessor( + changeArgs = ChangeArgs( + keys = Seq(UnqualifiedColumnName("id")), + sequencing = F.col("seq"), + storedAsScdType = ScdType.Type1 + ), + resolvedSequencingType = LongType + ) + + // Field names and dataTypes must match the input exactly, in the original order. + assert( + columnNamesAndDataTypes(processor.deduplicateMicrobatch(batch).schema) == + columnNamesAndDataTypes(schema)) + } + + test("deduplicateMicrobatch returns an empty DataFrame with preserved schema") { + val schema = new StructType() + .add("id", IntegerType) + .add("seq", LongType) + .add("value", StringType) + + val batch = microbatchOf(schema)() + + val processor = Scd1BatchProcessor( + changeArgs = ChangeArgs( + keys = Seq(UnqualifiedColumnName("id")), + sequencing = F.col("seq"), + storedAsScdType = ScdType.Type1 + ), + resolvedSequencingType = LongType + ) + + val result = processor.deduplicateMicrobatch(batch) + assert(result.collect().isEmpty) + assert(columnNamesAndDataTypes(result.schema) == columnNamesAndDataTypes(schema)) + } + + test("extendMicrobatchRowsWithCdcMetadata classifies each row as a delete or an upsert " + + "per deleteCondition") { + val schema = new StructType() + .add("id", IntegerType) + .add("seq", LongType) + .add("is_delete", BooleanType) + + val batch = microbatchOf(schema)( + Row(1, 10L, false), + Row(2, 20L, true), + Row(3, 30L, false), + Row(4, 40L, true) + ) + + val processor = Scd1BatchProcessor( + changeArgs = ChangeArgs( + keys = Seq(UnqualifiedColumnName("id")), + sequencing = F.col("seq"), + storedAsScdType = ScdType.Type1, + deleteCondition = Some(F.col("is_delete") === true) + ), + resolvedSequencingType = LongType + ) + + // Mutual-exclusivity invariant: each row's _cdc_metadata struct has exactly one of + // (deleteSequence, upsertSequence) non-null, and the non-null side carries the row's + // sequence value. + checkAnswer( + df = processor.extendMicrobatchRowsWithCdcMetadata(batch), + expectedAnswer = Seq( + Row(1, 10L, false, Row(null, 10L)), + Row(2, 20L, true, Row(20L, null)), + Row(3, 30L, false, Row(null, 30L)), + Row(4, 40L, true, Row(40L, null)) + ) + ) + } + + test("extendMicrobatchRowsWithCdcMetadata treats every row as an upsert " + + "when deleteCondition is None") { + val schema = new StructType() + .add("id", IntegerType) + .add("seq", LongType) + .add("value", StringType) + + val batch = microbatchOf(schema)( + Row(1, 10L, "a"), + Row(2, 20L, "b") + ) + + val processor = Scd1BatchProcessor( + changeArgs = ChangeArgs( + keys = Seq(UnqualifiedColumnName("id")), + sequencing = F.col("seq"), + storedAsScdType = ScdType.Type1, + deleteCondition = None + ), + resolvedSequencingType = LongType + ) + + checkAnswer( + df = processor.extendMicrobatchRowsWithCdcMetadata(batch), + expectedAnswer = Seq( + Row(1, 10L, "a", Row(null, 10L)), + Row(2, 20L, "b", Row(null, 20L)) + ) + ) + } + + test("extendMicrobatchRowsWithCdcMetadata appends CDC metadata as the last column") { + val schema = new StructType() + .add("id", IntegerType) + .add("seq", LongType) + .add("value", StringType) + + val batch = microbatchOf(schema)( + Row(1, 10L, "a") + ) + + val processor = Scd1BatchProcessor( + changeArgs = ChangeArgs( + keys = Seq(UnqualifiedColumnName("id")), + sequencing = F.col("seq"), + storedAsScdType = ScdType.Type1 + ), + resolvedSequencingType = LongType + ) + + val result = processor.extendMicrobatchRowsWithCdcMetadata(batch) + + // Original columns are preserved in their original order, with CDC metadata appended at + // the very end. + assert(result.schema.fieldNames.toSeq == + schema.fieldNames.toSeq :+ Scd1BatchProcessor.cdcMetadataColName) + } + + test("extendMicrobatchRowsWithCdcMetadata casts delete / upsert sequence fields to " + + "resolvedSequencingType") { + val schema = new StructType() + .add("id", IntegerType) + // Microbatch's sequencing column is IntegerType, but the flow's resolved sequencing type + // will be LongType. This should be upcasted in the projected CDC metadata column. + .add("seq", IntegerType) + .add("value", StringType) + + val batch = microbatchOf(schema)( + Row(1, 10, "a") + ) + + val processor = Scd1BatchProcessor( + changeArgs = ChangeArgs( + keys = Seq(UnqualifiedColumnName("id")), + sequencing = F.col("seq"), + storedAsScdType = ScdType.Type1 + ), + resolvedSequencingType = LongType + ) + + val resultDf = processor.extendMicrobatchRowsWithCdcMetadata(batch) + + val cdcMetadataDataType = + resultDf.schema(Scd1BatchProcessor.cdcMetadataColName).dataType.asInstanceOf[StructType] + assert(columnNamesAndDataTypes(cdcMetadataDataType) == Seq( + Scd1BatchProcessor.cdcDeleteSequenceFieldName -> LongType, + Scd1BatchProcessor.cdcUpsertSequenceFieldName -> LongType)) + + // The cast must also succeed at runtime: upsertSequence is materialized as a Long value, not + // an Int. + checkAnswer( + df = resultDf, + expectedAnswer = Row(1, 10, "a", Row(null, 10L)) + ) + } + + test("extendMicrobatchRowsWithCdcMetadata fails fast when the microbatch's sequencing column " + + "is incompatible with resolvedSequencingType") { + val schema = new StructType() + .add("id", IntegerType) + // Microbatch's sequencing column is a struct, whereas the flow's resolved sequencing type + // will be LongType. These are incompatible and should throw. + .add( + "seq", + new StructType() + .add("major", LongType) + .add("minor", LongType)) + + val batch = microbatchOf(schema)( + Row(1, Row(1L, 0L)) + ) + + val processor = Scd1BatchProcessor( + changeArgs = ChangeArgs( + keys = Seq(UnqualifiedColumnName("id")), + sequencing = F.col("seq"), + storedAsScdType = ScdType.Type1 + ), + resolvedSequencingType = LongType + ) + + val ex = intercept[AnalysisException] { + // .schema forces analysis of the underlying logical plan, surfacing the invalid cast. + processor.extendMicrobatchRowsWithCdcMetadata(batch).schema + } + assert(ex.getCondition == "DATATYPE_MISMATCH.CAST_WITHOUT_SUGGESTION") + } + + test("extendMicrobatchRowsWithCdcMetadata rejects a microbatch that already contains the " + + "reserved CDC metadata column") { + withSQLConf(SQLConf.CASE_SENSITIVE.key -> "true") { Review Comment: Reserved-name conflict is only tested with `CASE_SENSITIVE=true`. Consider adding `CASE_SENSITIVE=false` with a user column like `_CDC_METADATA` to exercise `validateCdcMetadataColumnNotPresent` + `resolver` under case-insensitive mode. ########## sql/pipelines/src/test/scala/org/apache/spark/sql/pipelines/autocdc/Scd1BatchProcessorSuite.scala: ########## @@ -0,0 +1,443 @@ +/* + * 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.pipelines.autocdc + +import org.apache.spark.SparkFunSuite +import org.apache.spark.sql.{functions => F, AnalysisException, Row} +import org.apache.spark.sql.classic.DataFrame +import org.apache.spark.sql.internal.SQLConf +import org.apache.spark.sql.test.SharedSparkSession +import org.apache.spark.sql.types._ + +class Scd1BatchProcessorSuite extends SparkFunSuite with SharedSparkSession { Review Comment: Nit (carried from #55969): Spark convention for suites using `checkAnswer` is `class Scd1BatchProcessorSuite extends QueryTest with SharedSparkSession` rather than `SparkFunSuite with SharedSparkSession`. ########## sql/pipelines/src/main/scala/org/apache/spark/sql/pipelines/autocdc/Scd1BatchProcessor.scala: ########## @@ -0,0 +1,161 @@ +/* + * 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.pipelines.autocdc + +import org.apache.spark.SparkException +import org.apache.spark.sql.{functions => F, AnalysisException} +import org.apache.spark.sql.Column +import org.apache.spark.sql.catalyst.util.QuotingUtils +import org.apache.spark.sql.classic.DataFrame +import org.apache.spark.sql.types.{DataType, StructField, StructType} +import org.apache.spark.util.ArrayImplicits._ + +/** + * Per-microbatch processor for SCD Type 1 AutoCDC flows, complying to the specified [[changeArgs]] + * configuration. + * + * @param changeArgs The CDC flow configuration. + * @param resolvedSequencingType The post-analysis [[DataType]] of the sequencing column, derived + * from the flow's resolved DataFrame at flow setup time. + */ +case class Scd1BatchProcessor( + changeArgs: ChangeArgs, + resolvedSequencingType: DataType) { + + /** + * Deduplicate the incoming CDC microbatch by key, keeping the most recent event per key + * as ordered by [[ChangeArgs.sequencing]]. + * + * For SCD1 we only care about the most recent (by sequence value) event per key. When + * multiple events share the same key and the same sequence value, the row selected is + * non-deterministic and undefined. + * + * The schema of the returned dataframe matches the schema of the microbatch exactly. + */ + def deduplicateMicrobatch(microbatchDf: DataFrame): DataFrame = { + // The `max_by` API can only return a single column, so pack/unpack the entire row into a + // temporary column before and after the `max_by` operation. + val winningRowCol = OutOfOrderCdcMergeUtils.tempColName("__winning_row") + + val allMicrobatchColumns = + microbatchDf.columns + .map(colName => F.col(QuotingUtils.quoteIdentifier(colName))) + .toImmutableArraySeq + + microbatchDf + .groupBy(changeArgs.keys.map(k => F.col(k.quoted)): _*) + .agg( + F.max_by(F.struct(allMicrobatchColumns: _*), changeArgs.sequencing) + .as(winningRowCol) + ) + .select(F.col(s"$winningRowCol.*")) + } + + /** + * Project the CDC metadata column onto the microbatch. + * + * The returned dataframe has all of the columns in the input microbatch + the CDC metadata + * column. + */ + def extendMicrobatchRowsWithCdcMetadata(microbatchDf: DataFrame): DataFrame = { + // Proactively validate the reserved CDC metadata column does not exist in the microbatch. + validateCdcMetadataColumnNotPresent(microbatchDf) + + val rowDeleteSequence: Column = changeArgs.deleteCondition match { + case Some(deleteCondition) => + F.when(deleteCondition, changeArgs.sequencing).otherwise(F.lit(null)) + case None => + F.lit(null) + } + + val rowUpsertSequence: Column = + // A row that is not a delete must be an upsert, these are mutually exclusive and a complete + // set of CDC event types. + F.when(rowDeleteSequence.isNull, changeArgs.sequencing).otherwise(F.lit(null)) + + microbatchDf.withColumn( + Scd1BatchProcessor.cdcMetadataColName, + Scd1BatchProcessor.constructCdcMetadataCol( + deleteSequence = rowDeleteSequence, + upsertSequence = rowUpsertSequence, + sequencingType = resolvedSequencingType + ) + ) + } + + private def validateCdcMetadataColumnNotPresent(microbatchDf: DataFrame): Unit = { + val sqlConf = microbatchDf.sparkSession.sessionState.conf + val resolver = sqlConf.resolver + + microbatchDf.schema.fieldNames + .find(resolver(_, Scd1BatchProcessor.cdcMetadataColName)) + .foreach { conflictingColumnName => + throw new AnalysisException( + errorClass = "AUTOCDC_RESERVED_COLUMN_NAME_CONFLICT", + messageParameters = Map( + "caseSensitivity" -> CaseSensitivityLabels.of(!sqlConf.caseSensitiveAnalysis), + "columnName" -> conflictingColumnName, + "schemaName" -> "microbatch", + "reservedColumnName" -> Scd1BatchProcessor.cdcMetadataColName + ) + ) + } + } +} + +object Scd1BatchProcessor { + private[autocdc] val cdcMetadataColName: String = "_cdc_metadata" + + private[autocdc] val cdcDeleteSequenceFieldName: String = "deleteSequence" Review Comment: Nit: the PR description uses `deleteVersion` / `upsertVersion`; the implementation uses `deleteSequence` / `upsertSequence`. Consider aligning the PR text (or a brief comment mapping to SPIP terms) so reviewers aren't confused. ########## sql/pipelines/src/main/scala/org/apache/spark/sql/pipelines/autocdc/Scd1BatchProcessor.scala: ########## @@ -0,0 +1,161 @@ +/* + * 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.pipelines.autocdc + +import org.apache.spark.SparkException +import org.apache.spark.sql.{functions => F, AnalysisException} +import org.apache.spark.sql.Column +import org.apache.spark.sql.catalyst.util.QuotingUtils +import org.apache.spark.sql.classic.DataFrame +import org.apache.spark.sql.types.{DataType, StructField, StructType} +import org.apache.spark.util.ArrayImplicits._ + +/** + * Per-microbatch processor for SCD Type 1 AutoCDC flows, complying to the specified [[changeArgs]] + * configuration. + * + * @param changeArgs The CDC flow configuration. + * @param resolvedSequencingType The post-analysis [[DataType]] of the sequencing column, derived + * from the flow's resolved DataFrame at flow setup time. + */ +case class Scd1BatchProcessor( + changeArgs: ChangeArgs, + resolvedSequencingType: DataType) { + + /** + * Deduplicate the incoming CDC microbatch by key, keeping the most recent event per key + * as ordered by [[ChangeArgs.sequencing]]. + * + * For SCD1 we only care about the most recent (by sequence value) event per key. When + * multiple events share the same key and the same sequence value, the row selected is + * non-deterministic and undefined. + * + * The schema of the returned dataframe matches the schema of the microbatch exactly. + */ + def deduplicateMicrobatch(microbatchDf: DataFrame): DataFrame = { + // The `max_by` API can only return a single column, so pack/unpack the entire row into a + // temporary column before and after the `max_by` operation. + val winningRowCol = OutOfOrderCdcMergeUtils.tempColName("__winning_row") + + val allMicrobatchColumns = + microbatchDf.columns + .map(colName => F.col(QuotingUtils.quoteIdentifier(colName))) + .toImmutableArraySeq + + microbatchDf + .groupBy(changeArgs.keys.map(k => F.col(k.quoted)): _*) + .agg( + F.max_by(F.struct(allMicrobatchColumns: _*), changeArgs.sequencing) + .as(winningRowCol) + ) + .select(F.col(s"$winningRowCol.*")) + } + + /** + * Project the CDC metadata column onto the microbatch. + * + * The returned dataframe has all of the columns in the input microbatch + the CDC metadata + * column. + */ + def extendMicrobatchRowsWithCdcMetadata(microbatchDf: DataFrame): DataFrame = { Review Comment: Tests exercise `deduplicateMicrobatch` and `extendMicrobatchRowsWithCdcMetadata` separately. Consider a short integration test on the **deduped** microbatch (e.g. two rows same key, winner is delete) to show metadata reflects the winning row only — matching real foreachBatch ordering (`dedup` then `extend`). ########## sql/pipelines/src/main/scala/org/apache/spark/sql/pipelines/autocdc/Scd1BatchProcessor.scala: ########## @@ -0,0 +1,161 @@ +/* + * 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.pipelines.autocdc + +import org.apache.spark.SparkException +import org.apache.spark.sql.{functions => F, AnalysisException} +import org.apache.spark.sql.Column +import org.apache.spark.sql.catalyst.util.QuotingUtils +import org.apache.spark.sql.classic.DataFrame +import org.apache.spark.sql.types.{DataType, StructField, StructType} +import org.apache.spark.util.ArrayImplicits._ + +/** + * Per-microbatch processor for SCD Type 1 AutoCDC flows, complying to the specified [[changeArgs]] + * configuration. + * + * @param changeArgs The CDC flow configuration. + * @param resolvedSequencingType The post-analysis [[DataType]] of the sequencing column, derived + * from the flow's resolved DataFrame at flow setup time. + */ +case class Scd1BatchProcessor( + changeArgs: ChangeArgs, + resolvedSequencingType: DataType) { + + /** + * Deduplicate the incoming CDC microbatch by key, keeping the most recent event per key + * as ordered by [[ChangeArgs.sequencing]]. + * + * For SCD1 we only care about the most recent (by sequence value) event per key. When + * multiple events share the same key and the same sequence value, the row selected is + * non-deterministic and undefined. + * + * The schema of the returned dataframe matches the schema of the microbatch exactly. + */ + def deduplicateMicrobatch(microbatchDf: DataFrame): DataFrame = { + // The `max_by` API can only return a single column, so pack/unpack the entire row into a + // temporary column before and after the `max_by` operation. + val winningRowCol = OutOfOrderCdcMergeUtils.tempColName("__winning_row") + + val allMicrobatchColumns = + microbatchDf.columns + .map(colName => F.col(QuotingUtils.quoteIdentifier(colName))) + .toImmutableArraySeq + + microbatchDf + .groupBy(changeArgs.keys.map(k => F.col(k.quoted)): _*) + .agg( + F.max_by(F.struct(allMicrobatchColumns: _*), changeArgs.sequencing) + .as(winningRowCol) + ) + .select(F.col(s"$winningRowCol.*")) + } + + /** + * Project the CDC metadata column onto the microbatch. Review Comment: Worth noting in scaladoc: this should run **before** `columnSelection` drops columns required by `deleteCondition`, since classification is computed from `deleteCondition` + `sequencing` on the current microbatch schema. Helps future wiring PRs get call order right. -- 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]
