gaborgsomogyi commented on a change in pull request #24922: [SPARK-28120][SS] Rocksdb state storage implementation URL: https://github.com/apache/spark/pull/24922#discussion_r316178811
########## File path: sql/core/src/main/scala/org/apache/spark/sql/execution/streaming/state/RocksDbStateStoreProvider.scala ########## @@ -0,0 +1,642 @@ +/* + * 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.io._ +import java.util +import java.util.Locale + +import scala.collection.JavaConverters._ +import scala.util.control.NonFatal + +import org.apache.commons.io.FileUtils +import org.apache.hadoop.conf.Configuration +import org.apache.hadoop.fs._ + +import org.apache.spark.{SparkConf, SparkEnv} +import org.apache.spark.internal.Logging +import org.apache.spark.io.FileUtility +import org.apache.spark.sql.catalyst.expressions.UnsafeRow +import org.apache.spark.sql.execution.streaming.CheckpointFileManager +import org.apache.spark.sql.types.StructType +import org.apache.spark.util.Utils + +private[sql] class RocksDbStateStoreProvider extends StateStoreProvider with Logging { + + /* Internal fields and methods */ + @volatile private var stateStoreId_ : StateStoreId = _ + @volatile private var keySchema: StructType = _ + @volatile private var valueSchema: StructType = _ + @volatile private var storeConf: StateStoreConf = _ + @volatile private var hadoopConf: Configuration = _ + @volatile private var numberOfVersionsToRetain: Int = _ + @volatile private var localDirectory: String = _ + + /* + * Additional configurations related to rocksDb. This will capture all configs in + * SQLConf that start with `spark.sql.streaming.stateStore.rocksDb` + */ + @volatile private var rocksDbConf: Map[String, String] = Map.empty[String, String] + + private lazy val baseDir: Path = stateStoreId.storeCheckpointLocation() + private lazy val fm = CheckpointFileManager.create(baseDir, hadoopConf) + private lazy val sparkConf = Option(SparkEnv.get).map(_.conf).getOrElse(new SparkConf) + + private case class StoreFile(version: Long, path: Path, isSnapshot: Boolean) + + import WALUtils._ + + /** Implementation of [[StateStore]] API which is backed by RocksDB and HDFS */ + class RocksDbStateStore(val version: Long) extends StateStore with Logging { + + /** Trait and classes representing the internal state of the store */ + trait STATE + + case object LOADED extends STATE + + case object UPDATING extends STATE + + case object COMMITTED extends STATE + + case object ABORTED extends STATE + + private val newVersion = version + 1 + @volatile private var state: STATE = LOADED + private val finalDeltaFile: Path = deltaFile(baseDir, newVersion) + private lazy val deltaFileStream = fm.createAtomic(finalDeltaFile, overwriteIfPossible = true) + private lazy val compressedStream = compressStream(deltaFileStream, sparkConf) + + override def id: StateStoreId = RocksDbStateStoreProvider.this.stateStoreId + + var rocksDbWriteInstance: OptimisticTransactionDbInstance = null + + /* + * numEntriesInDb and bytesUsedByDb are estimated value + * due to the nature of RocksDB implementation. + * see https://github.com/facebook/rocksdb/wiki/RocksDB-FAQ for more details + */ + var numEntriesInDb: Long = 0L + var bytesUsedByDb: Long = 0L + + private def initTransaction(): Unit = { + if (state == LOADED && rocksDbWriteInstance == null) { + logDebug(s"Creating Transactional DB for batch $version") + rocksDbWriteInstance = new OptimisticTransactionDbInstance( + keySchema, + valueSchema, + newVersion.toString, + rocksDbConf) + rocksDbWriteInstance.open(rocksDbPath) + state = UPDATING + rocksDbWriteInstance.startTransactions() + } + } + + override def get(key: UnsafeRow): UnsafeRow = { + initTransaction() + rocksDbWriteInstance.get(key) + } + + override def put(key: UnsafeRow, value: UnsafeRow): Unit = { + initTransaction() + verify(state == UPDATING, s"Cannot put after already committed or aborted") + val keyCopy = key.copy() + val valueCopy = value.copy() + rocksDbWriteInstance.put(keyCopy, valueCopy) + writeUpdateToDeltaFile(compressedStream, keyCopy, valueCopy) + } + + override def remove(key: UnsafeRow): Unit = { + initTransaction() + verify(state == UPDATING, "Cannot remove after already committed or aborted") + rocksDbWriteInstance.remove(key) + writeRemoveToDeltaFile(compressedStream, key) + } + + override def getRange( + start: Option[UnsafeRow], + end: Option[UnsafeRow]): Iterator[UnsafeRowPair] = { + verify(state == UPDATING, "Cannot getRange after already committed or aborted") + iterator() + } + + /** Commit all the updates that have been made to the store, and return the new version. */ + override def commit(): Long = { + initTransaction() + verify(state == UPDATING, s"Cannot commit after already committed or aborted") + try { + synchronized { + rocksDbWriteInstance.commit(Some(getCheckpointPath(newVersion))) + finalizeDeltaFile(compressedStream) + } + state = COMMITTED + numEntriesInDb = rocksDbWriteInstance.getApproxEntriesInDb() + bytesUsedByDb = numEntriesInDb * (keySchema.defaultSize + valueSchema.defaultSize) + newVersion + } catch { + case NonFatal(e) => + throw new IllegalStateException(s"Error committing version $newVersion into $this", e) + } finally { + storeMap.remove(version) + close() + } + } + + /* + * Abort all the updates made on this store. This store will not be usable any more. + */ + override def abort(): Unit = { + // This if statement is to ensure that files are deleted only if there are changes to the + // StateStore. We have two StateStores for each task, one which is used only for reading, and + // the other used for read+write. We don't want the read-only to delete state files. + try { + if (state == UPDATING) { + state = ABORTED + synchronized { + rocksDbWriteInstance.abort() + cancelDeltaFile(compressedStream, deltaFileStream) + } + logInfo(s"Aborted version $newVersion for $this") + } else { + state = ABORTED + } + } catch { + case NonFatal(e) => + throw new IllegalStateException(s"Error aborting version $newVersion into $this", e) + } finally { + storeMap.remove(version) + close() + } + } + + def close(): Unit = { + if (rocksDbWriteInstance != null) { + rocksDbWriteInstance.close() + rocksDbWriteInstance = null + } + } + + /* + * Get an iterator of all the store data. + * This can be called only after committing all the updates made in the current thread. + */ + override def iterator(): Iterator[UnsafeRowPair] = { + state match { + case UPDATING => + logDebug("state = updating using transaction DB") + // We need to use current db to read uncommitted transactions + rocksDbWriteInstance.iterator(closeDbOnCompletion = false) + + case LOADED | ABORTED => + // use check-pointed db for previous version + logDebug(s"state = loaded/aborted using check-pointed DB with version $version") + if (version == 0) { + Iterator.empty + } else { + val path = getCheckpointPath(version) + val r: RocksDbInstance = + new RocksDbInstance(keySchema, valueSchema, version.toString, rocksDbConf) + r.open(path, readOnly = true) + r.iterator(closeDbOnCompletion = true) + } + case COMMITTED => + logDebug(s"state = committed using check-pointed DB with version $newVersion") + // use check-pointed db for current updated version + val path = getCheckpointPath(newVersion) + val r: RocksDbInstance = + new RocksDbInstance(keySchema, valueSchema, newVersion.toString, rocksDbConf) + r.open(path, readOnly = true) + r.iterator(closeDbOnCompletion = true) + + case _ => Iterator.empty + } + } + + override def metrics: StateStoreMetrics = { + val metricsFromProvider: Map[String, Long] = getMetricsForProvider() + val customMetrics = metricsFromProvider.flatMap { + case (name, value) => + // just allow searching from list cause the list is small enough + supportedCustomMetrics.find(_.name == name).map(_ -> value) + } + StateStoreMetrics(Math.max(numEntriesInDb, 0), Math.max(bytesUsedByDb, 0), customMetrics) + } + + /* + * Whether all updates have been committed + */ + override def hasCommitted: Boolean = { + state == COMMITTED + } + + override def toString(): String = { + s"RocksDbStateStore[id=(op=${id.operatorId},part=${id.partitionId}),dir=$baseDir]" + } + + } + + /* + * Initialize the provider with more contextual information from the SQL operator. + * This method will be called first after creating an instance of the StateStoreProvider by + * reflection. + * + * @param stateStoreId Id of the versioned StateStores that this provider will generate + * @param keySchema Schema of keys to be stored + * @param valueSchema Schema of value to be stored + * @param keyIndexOrdinal Optional column (represent as the ordinal of the field in keySchema) by + * which the StateStore implementation could index the data. + * @param storeConfs Configurations used by the StateStores + * @param hadoopConf Hadoop configuration that could be used by StateStore + * to save state data + */ + override def init( + stateStoreId: StateStoreId, + keySchema: StructType, + valueSchema: StructType, + keyIndexOrdinal: Option[Int], // for sorting the data by their keys + storeConfs: StateStoreConf, + hadoopConf: Configuration): Unit = { + this.stateStoreId_ = stateStoreId + this.keySchema = keySchema + this.valueSchema = valueSchema + this.storeConf = storeConfs + this.hadoopConf = hadoopConf + this.numberOfVersionsToRetain = storeConfs.maxVersionsToRetainInMemory + fm.mkdirs(baseDir) + this.rocksDbConf = storeConf.confs + .filter(_._1.startsWith("spark.sql.streaming.stateStore.rocksDb")) + .map { + case (k, v) => (k.toLowerCase(Locale.ROOT), v) + } + this.localDirectory = this.rocksDbConf + .getOrElse( + "spark.sql.streaming.stateStore.rocksDb.localDirectory".toLowerCase(Locale.ROOT), + Utils.createTempDir().getAbsoluteFile.toString) + } + + /* + * Return the id of the StateStores this provider will generate. + * Should be the same as the one passed in init(). + */ + override def stateStoreId: StateStoreId = stateStoreId_ + + /* + * Called when the provider instance is unloaded from the executor + */ + override def close(): Unit = { + storeMap.values.asScala.foreach(_.close) + storeMap.clear() + } + + private val storeMap = new util.HashMap[Long, RocksDbStateStore]() + + /* + * Optional custom metrics that the implementation may want to report. + * + * @note The StateStore objects created by this provider must report the same custom metrics + * (specifically, same names) through `StateStore.metrics`. + */ + // TODO + override def supportedCustomMetrics: Seq[StateStoreCustomMetric] = { + Nil + } + + override def toString(): String = { + s"RocksDbStateStoreProvider[" + + s"id = (op=${stateStoreId.operatorId},part=${stateStoreId.partitionId}),dir = $baseDir]" + } + + def getMetricsForProvider(): Map[String, Long] = synchronized { + Map.empty[String, Long] + } + + /* + * Return an instance of [[StateStore]] representing state data of the given version + */ + override def getStore(version: Long): StateStore = synchronized { + logInfo(s"get Store for version $version") + require(version >= 0, "Version cannot be less than 0") + if (storeMap.containsKey(version)) { + storeMap.get(version) + } else { + val store = createStore(version) + storeMap.put(version, store) + store + } + } + + private def createStore(version: Long): RocksDbStateStore = { + val newStore = new RocksDbStateStore(version) + if (version > 0) { + // load the data into the rocksDB + logInfo( + s"Loading state into the db for $version and partition ${stateStoreId_.partitionId}") + loadIntoRocksDB(version) + } + newStore + } + + private def loadIntoRocksDB(version: Long): Unit = { + /* + 1. Get last available/committed Rocksdb version in local folder + 2. If last committed version = version, we already have loaded rocksdb state. + 3. If last committed version = version - 1, + we have to apply delta for version in the existing rocksdb + 4. Otherwise we have to recreate a new rocksDB store by using Snapshots/Delta + */ + val (_, elapsedMs) = Utils.timeTakenMs { + var lastAvailableVersion = getLastCommittedVersion() + if (lastAvailableVersion == -1L || lastAvailableVersion <= version - 2) { + // Destroy existing DB so that we can reconstruct it using snapshot and delta files + RocksDbInstance.destroyDB(rocksDbPath) + var lastAvailableSnapShotVersion: Long = version + 1 + // load from snapshot + var found = false + while (!found && lastAvailableSnapShotVersion > 0) { + try { + lastAvailableSnapShotVersion = lastAvailableSnapShotVersion - 1 + found = loadSnapshotFile(lastAvailableSnapShotVersion) + logDebug( + s"Snapshot for version $lastAvailableSnapShotVersion " + + "and partition ${stateStoreId_.partitionId}: found = $found") + } catch { + case e: Exception => + logError(s"$e while reading snapshot file") + throw e + } + } + lastAvailableVersion = lastAvailableSnapShotVersion + } + if (lastAvailableVersion < version) { + applyDelta(version, lastAvailableVersion) + } + } + logInfo( + s"Loading state for $version and partition ${stateStoreId_.partitionId} took $elapsedMs ms.") + } + + private def getLastCommittedVersion(): Long = { + val f = new File(rocksDbPath, RocksDbInstance.COMMIT_FILE_NAME) + if (f.exists()) { + try { + val fileContents = scala.io.Source.fromFile(f.getAbsolutePath).getLines.mkString + return fileContents.toLong + } catch { + case e: Exception => + logWarning("Exception while reading committed file") + } + } + return -1L + } + + private def loadSnapshotFile(version: Long): Boolean = { + val fileToRead = snapshotFile(baseDir, version) + if (version == 0 || !fm.exists(fileToRead)) { + return false + } + val versionTempPath = getTempPath(version) + val tmpLocDir: File = new File(versionTempPath) + val tmpLocFile: File = new File(s"${versionTempPath}.tar") + try { + logInfo(s"Will download $fileToRead at location ${tmpLocFile.toString()}") + if (downloadFile(fm, fileToRead, new Path(tmpLocFile.getAbsolutePath), sparkConf)) { + FileUtility.extractTarFile(tmpLocFile.getAbsolutePath, versionTempPath) + if (!tmpLocDir.list().exists(_.endsWith(".sst"))) { + logWarning("Snapshot files are corrupted") + throw new IOException( + s"Error reading snapshot file $fileToRead of $this:" + + s" No SST files found") + } + FileUtils.moveDirectory(tmpLocDir, new File(rocksDbPath)) + true + } else { + false + } + } catch { + case e: Exception => + logError(s"Exception while loading snapshot file $e") + throw e + } finally { + if (tmpLocFile.exists()) { + tmpLocFile.delete() + } + FileUtils.deleteDirectory(tmpLocDir) + } + } + + private def applyDelta(version: Long, lastAvailableVersion: Long): Unit = { + var rocksDbWriteInstance: OptimisticTransactionDbInstance = null + try { + rocksDbWriteInstance = + new OptimisticTransactionDbInstance(keySchema, valueSchema, version.toString, rocksDbConf) + rocksDbWriteInstance.open(rocksDbPath) + rocksDbWriteInstance.startTransactions() + // Load all the deltas from the version after the last available + // one up to the target version. + // The last available version is the one with a full snapshot, so it doesn't need deltas. + for (deltaVersion <- (lastAvailableVersion + 1) to version) { + val fileToRead = deltaFile(baseDir, deltaVersion) + updateFromDeltaFile( + fm, + fileToRead, + keySchema, + valueSchema, + rocksDbWriteInstance, + sparkConf) + logInfo(s"Read delta file for version $version of $this from $fileToRead") + } + rocksDbWriteInstance.commit(Some(getCheckpointPath(version))) + } catch { + case e: Exception => + logError(s"Exception while loading state ${e.getMessage}") + if (rocksDbWriteInstance != null) { + rocksDbWriteInstance.abort() + } + throw e + } finally { + if (rocksDbWriteInstance != null) { + rocksDbWriteInstance.close() + } + } + } + + /** Optional method for providers to allow for background maintenance (e.g. compactions) */ + override def doMaintenance(): Unit = { + try { + val (files: Seq[WALUtils.StoreFile], e1) = Utils.timeTakenMs(fetchFiles(fm, baseDir)) + logDebug(s"fetchFiles() took $e1 ms.") + doSnapshot(files) + cleanup(files) + cleanRocksDBCheckpoints(files) + } catch { + case NonFatal(e) => + logWarning(s"Error performing snapshot and cleaning up $this") + } + } + + private def doSnapshot(files: Seq[WALUtils.StoreFile]): Unit = { + if (files.nonEmpty) { + val lastVersion = files.last.version + val deltaFilesForLastVersion = + filesForVersion(files, lastVersion).filter(_.isSnapshot == false) + if (deltaFilesForLastVersion.size > storeConf.minDeltasForSnapshot) { + val dbPath = getCheckpointPath(lastVersion) + val snapShotFileName = s"{getTempPath(lastVersion)}.snapshot" + val f = new File(snapShotFileName) + try { + val (_, t1) = Utils.timeTakenMs { + FileUtility.createTarFile(dbPath, snapShotFileName) + val targetFile = snapshotFile(baseDir, lastVersion) + uploadFile(fm, new Path(snapShotFileName), targetFile, sparkConf) + } + logInfo(s"Creating snapshot file for ${stateStoreId_.partitionId} took $t1 ms.") + } catch { + case e: Exception => + logError(s"Exception while creating snapshot $e") + throw e + } finally { + f.delete() // delete the tarball + } + } + } + } + + /* + * Clean up old snapshots and delta files that are not needed any more. It ensures that last + * few versions of the store can be recovered from the files, so re-executed RDD operations + * can re-apply updates on the past versions of the store. + */ + private[state] def cleanup(files: Seq[WALUtils.StoreFile]): Unit = { + try { + if (files.nonEmpty) { + val earliestVersionToRetain = files.last.version - storeConf.minVersionsToRetain + if (earliestVersionToRetain > 0) { + val earliestFileToRetain = filesForVersion(files, earliestVersionToRetain).head + val filesToDelete = files.filter(_.version < earliestFileToRetain.version) + val (_, e2) = Utils.timeTakenMs { + filesToDelete.foreach { f => + fm.delete(f.path) + val file = new File(rocksDbPath, f.version.toString) + if (file.exists()) { + file.delete() + } + } + } + logDebug(s"deleting files took $e2 ms.") + logInfo( + s"Deleted files older than ${earliestFileToRetain.version} for $this: " + + filesToDelete.mkString(", ")) + } + } + } catch { + case NonFatal(e) => + logWarning(s"Error cleaning up files for $this", e) + } + } + + private def cleanRocksDBCheckpoints(files: Seq[WALUtils.StoreFile]): Unit = { + try { + val (_, e2) = Utils.timeTakenMs { + if (files.nonEmpty) { + val earliestVersionToRetain = files.last.version - storeConf.minVersionsToRetain + if (earliestVersionToRetain > 0) { + new File(getCheckpointPath(earliestVersionToRetain)).getParentFile + .listFiles(new FileFilter { + def accept(f: File): Boolean = { + try { + f.getName.toLong < earliestVersionToRetain + } catch { + case _: NumberFormatException => false + } + } + }) + .foreach(p => RocksDbInstance.destroyDB(p.getAbsolutePath)) + logInfo( + s"Deleted rocksDB checkpoints older than ${earliestVersionToRetain} for $this: ") + } + } + } + logDebug(s"deleting rocksDB checkpoints took $e2 ms.") + } catch { + case NonFatal(e) => logWarning(s"Error cleaning up files for $this", e) + } + } + + // Used only for unit tests + private[sql] def latestIterator(): Iterator[UnsafeRowPair] = synchronized { + val versionsInFiles = fetchFiles(fm, baseDir).map(_.version).toSet + if (versionsInFiles.nonEmpty) { + val maxVersion = versionsInFiles.max + if (maxVersion > 0) { + loadIntoRocksDB(maxVersion) + val r: RocksDbInstance = + new RocksDbInstance(keySchema, valueSchema, maxVersion.toString, rocksDbConf) + try { + r.open(rocksDbPath, readOnly = true) + return r.iterator(false) + } catch { + case e: Exception => + logWarning(s"Exception ${e.getMessage} while getting latest Iterator") + } + } + } + Iterator.empty + } + + private[sql] def getLocalDirectory: String = localDirectory Review comment: As I see it's only used from tests. If the tests are setting `spark.sql.streaming.stateStore.rocksDb.localDirectory` then no need this, 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. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org
