micheal-o commented on code in PR #53703: URL: https://github.com/apache/spark/pull/53703#discussion_r2666821684
########## sql/core/src/main/scala/org/apache/spark/sql/execution/streaming/state/StateRewriter.scala: ########## @@ -0,0 +1,373 @@ +/* + * 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.execution.streaming.state + +import java.util.UUID + +import org.apache.hadoop.conf.Configuration +import org.apache.hadoop.fs.Path + +import org.apache.spark.{SparkIllegalStateException, TaskContext} +import org.apache.spark.internal.Logging +import org.apache.spark.internal.LogKeys._ +import org.apache.spark.sql.DataFrame +import org.apache.spark.sql.SparkSession +import org.apache.spark.sql.execution.datasources.v2.state.StateSourceOptions +import org.apache.spark.sql.execution.datasources.v2.state.metadata.StateMetadataPartitionReader +import org.apache.spark.sql.execution.streaming.checkpointing.OffsetSeqMetadata +import org.apache.spark.sql.execution.streaming.operators.stateful.StatefulOperatorsUtils +import org.apache.spark.sql.execution.streaming.operators.stateful.transformwithstate.{StateVariableType, TransformWithStateOperatorProperties, TransformWithStateVariableInfo} +import org.apache.spark.sql.execution.streaming.runtime.{StreamingCheckpointConstants, StreamingQueryCheckpointMetadata} +import org.apache.spark.sql.execution.streaming.state.{StatePartitionAllColumnFamiliesWriter, StateSchemaCompatibilityChecker} +import org.apache.spark.sql.internal.SQLConf +import org.apache.spark.util.SerializableConfiguration + +/** + * State Rewriter is used to rewrite the state stores for a stateful streaming query. + * It reads state from a checkpoint location, optionally applies transformation to the state, + * and then writes the state back to a (possibly different) checkpoint location for a new batch ID. + * + * Example use case is for offline state repartitioning. + * Can also be used to support other use cases. + * + * @param sparkSession The active Spark session. + * @param readBatchId The batch ID for reading state. + * @param writeBatchId The batch ID to which the (transformed) state will be written. + * @param resolvedCheckpointLocation The resolved checkpoint path where state will be written. + * @param hadoopConf Hadoop configuration for file system operations. + * @param readResolvedCheckpointLocation Optional separate checkpoint location to read state from. + * If None, reads from resolvedCheckpointLocation. + * @param transformFunc Optional transformation function applied to each operator's state + * DataFrame. If None, state is written as-is. + * @param writeCheckpointMetadata Optional checkpoint metadata for the resolvedCheckpointLocation. + * If None, will create a new one for resolvedCheckpointLocation. + * Helps us to reuse already cached checkpoint log entries, + * instead of starting from scratch. + */ +class StateRewriter( + sparkSession: SparkSession, + readBatchId: Long, + writeBatchId: Long, + resolvedCheckpointLocation: String, + hadoopConf: Configuration, + readResolvedCheckpointLocation: Option[String] = None, + transformFunc: Option[DataFrame => DataFrame] = None, + writeCheckpointMetadata: Option[StreamingQueryCheckpointMetadata] = None +) extends Logging { + require(readResolvedCheckpointLocation.isDefined || readBatchId < writeBatchId, + s"Read batch id $readBatchId must be less than write batch id $writeBatchId " + + "when reading and writing to the same checkpoint location") + + // If a different location was specified for reading state, use it. + // Else, use same location for reading and writing state. + private val checkpointLocationForRead = + readResolvedCheckpointLocation.getOrElse(resolvedCheckpointLocation) + private val stateRootLocation = new Path( + resolvedCheckpointLocation, StreamingCheckpointConstants.DIR_NAME_STATE).toString + + def run(): Unit = { + logInfo(log"Starting state rewrite for " + + log"checkpointLocation=${MDC(CHECKPOINT_LOCATION, resolvedCheckpointLocation)}, " + + log"readCheckpointLocation=" + + log"${MDC(CHECKPOINT_LOCATION, readResolvedCheckpointLocation.getOrElse(""))}, " + + log"readBatchId=${MDC(BATCH_ID, readBatchId)}, " + + log"writeBatchId=${MDC(BATCH_ID, writeBatchId)}") + + try { + val stateMetadataReader = new StateMetadataPartitionReader( + resolvedCheckpointLocation, + new SerializableConfiguration(hadoopConf), + readBatchId) + + val allOperatorsMetadata = stateMetadataReader.allOperatorStateMetadata + if (allOperatorsMetadata.isEmpty) { + // Could be the query is stateless or ran on older spark version without op metadata + throw StateRewriterErrors.missingOperatorMetadataError( + resolvedCheckpointLocation, readBatchId) + } + + // Use the same conf in the offset log to create the store conf, + // to make sure the state is written with the right conf. + val (storeConf, sqlConf) = createConfsFromOffsetLog() + // SQLConf doesn't serialize properly (reader becomes null), so extract as Map + val sqlConfEntries: Map[String, String] = sqlConf.getAllConfs + + // A Hadoop Configuration can be about 10 KB, which is pretty big, so broadcast it + val hadoopConfBroadcast = + SerializableConfiguration.broadcast(sparkSession.sparkContext, hadoopConf) + + // Do rewrite for each operator + // We can potentially parallelize this, but for now, do sequentially + allOperatorsMetadata.foreach { opMetadata => + val stateStoresMetadata = opMetadata.stateStoresMetadata + assert(!stateStoresMetadata.isEmpty, + s"Operator ${opMetadata.operatorInfo.operatorName} has no state stores") + + val storeToSchemaFilesMap = getStoreToSchemaFilesMap(opMetadata) + val stateVarsIfTws = getStateVariablesIfTWS(opMetadata) + + // Rewrite each state store of the operator + stateStoresMetadata.foreach { stateStoreMetadata => + // Read state + val stateDf = sparkSession.read + .format("statestore") + .option(StateSourceOptions.PATH, checkpointLocationForRead) + .option(StateSourceOptions.BATCH_ID, readBatchId) + .option(StateSourceOptions.OPERATOR_ID, opMetadata.operatorInfo.operatorId) + .option(StateSourceOptions.STORE_NAME, stateStoreMetadata.storeName) + .option(StateSourceOptions.INTERNAL_ONLY_READ_ALL_COLUMN_FAMILIES, "true") + .load() + + // Run the caller state transformation func if provided + // Otherwise, use the state as is + val updatedStateDf = transformFunc.map(func => func(stateDf)).getOrElse(stateDf) Review Comment: That will be specified/set by the repartition runner when it calls the state rewriter. Don't want to over couple the StateRewriter to repartition, to make it reusable for other use cases. -- 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]
