JingsongLi commented on code in PR #8214: URL: https://github.com/apache/paimon/pull/8214#discussion_r3445282858
########## paimon-spark/paimon-spark-common/src/main/scala/org/apache/paimon/spark/write/PaimonV2DataEvolutionDataWriterBase.scala: ########## @@ -0,0 +1,199 @@ +/* + * 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.paimon.spark.write + +import org.apache.paimon.catalog.CatalogContext +import org.apache.paimon.data.{BinaryRow, InternalRow => PaimonInternalRow} +import org.apache.paimon.io.{CompactIncrement, DataIncrement} +import org.apache.paimon.operation.AbstractFileStoreWrite +import org.apache.paimon.schema.TableSchema +import org.apache.paimon.spark.SparkInternalRowWrapper +import org.apache.paimon.spark.schema.PaimonMetadataColumn +import org.apache.paimon.table.sink.{BatchWriteBuilder, CommitMessage, CommitMessageImpl, RowPartitionKeyExtractor, TableWriteImpl} +import org.apache.paimon.utils.RecordWriter + +import org.apache.spark.internal.Logging +import org.apache.spark.sql.catalyst.InternalRow + +import java.util.Collections + +import scala.collection.JavaConverters._ +import scala.collection.mutable.ListBuffer + +/** + * V2 [[DataWrite]] for copy-on-write UPDATE on data-evolution tables. + * + * Data-evolution rows derive `_ROW_ID` from the file's `firstRowId` plus the row position, so + * rewritten files must hold contiguous row-id runs and carry a `firstRowId` themselves instead of + * writing physical row-tracking columns (files with a physical `_ROW_ID` column are not assigned a + * `firstRowId` at commit and cannot be grouped by the data-evolution read path). The surviving rows + * arrive range-distributed and sorted by `_ROW_ID` (see + * [[PaimonWriteRequirement.dataEvolutionCopyOnWrite]]); this writer starts a new file run whenever + * the row id is not consecutive or the partition changes, and assigns each produced file the run's + * row id offset, mirroring `DataEvolutionTableDataWrite.PerFileWriter` and the commit skip rule in + * `RowTrackingCommitUtils` (files with a non-null `firstRowId` keep it). Sequence numbers are not + * preserved: the new files keep `minSequenceNumber == 0` and the commit stamps them with the new + * snapshot id, the same as any data-evolution rewrite. + */ +abstract class PaimonV2DataEvolutionDataWriterBase( + writeBuilder: BatchWriteBuilder, + tableSchema: TableSchema, + writeSchema: org.apache.spark.sql.types.StructType, + dataSchema: org.apache.spark.sql.types.StructType, + metadataSchema: org.apache.spark.sql.types.StructType, + catalogContext: CatalogContext) + extends InnerTableV2DataWrite + with Logging { + + private val rowIdIndex = metadataSchema.fieldIndex(PaimonMetadataColumn.ROW_ID_COLUMN) + + private val partitionExtractor = new RowPartitionKeyExtractor(tableSchema) + + private val reusableWrapper = + new SparkInternalRowWrapper(writeSchema, writeSchema.fields.length, dataSchema, catalogContext) + + private var currentRun: RunWriter = _ + private val commitMessages = ListBuffer[CommitMessage]() + + protected def writeWithMetadata(metadata: InternalRow, data: InternalRow): Unit = { + assert( + metadata != null && !metadata.isNullAt(rowIdIndex), + "Data-evolution copy-on-write requires a non-null _ROW_ID for every written row.") + val rowId = metadata.getLong(rowIdIndex) + val paimonRow = reusableWrapper.replace(data) + val partition = partitionExtractor.partition(paimonRow) + if (currentRun == null || !currentRun.accepts(rowId, partition)) { + finishCurrentRun() + currentRun = new RunWriter(rowId, partition.copy()) + } + currentRun.write(paimonRow) + } + + override def write(record: InternalRow): Unit = { + throw new UnsupportedOperationException( + "Data-evolution V2 copy-on-write only rewrites existing rows; " + + "rows without row-tracking metadata are not supported.") + } + + override def commitImpl(): Seq[CommitMessage] = { + finishCurrentRun() + commitMessages.toSeq + } + + override def abort(): Unit = close() + + override def close(): Unit = { + if (currentRun != null) { + currentRun.closeQuietly() + currentRun = null + } + } + + private def finishCurrentRun(): Unit = { + if (currentRun != null) { + commitMessages ++= currentRun.finish() + currentRun = null + } + } + + /** Writes one contiguous row-id run; may roll into several files, all stay range-adjacent. */ + private class RunWriter(startRowId: Long, partition: BinaryRow) { + + private var nextRowId = startRowId + + // The run writer must only append the rewritten rows and never restore the bucket's + // existing files. The bucket-unaware append write already enforces this + // (AppendFileStoreWrite pins ignorePreviousFiles=true and uses NoopCompactManager), but + // declare it explicitly rather than depend on which FileStoreWrite subclass we got. + private val recordWriter: RecordWriter[PaimonInternalRow] = writeBuilder + .newWrite() + .asInstanceOf[TableWriteImpl[PaimonInternalRow]] + .withIgnorePreviousFiles(true) + .getWrite + .asInstanceOf[AbstractFileStoreWrite[PaimonInternalRow]] + .createWriter(partition, 0) Review Comment: This path skips the IOManager setup that the normal V2 writer does before creating append writers. If a data-evolution table has `write-buffer-for-append=true` (with the default spillable buffer), `AppendOnlyWriter` builds an `ExternalBuffer` with a null IOManager; once the buffer spills, it will fail at `ioManager.createChannel()`. Please pass an IOManager into this `TableWriteImpl` (and close it with the writer), or otherwise disable the buffered append path here. -- 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]
