viirya commented on code in PR #58978: URL: https://github.com/apache/spark/pull/58978#discussion_r4235886225
########## sql/core/src/main/scala/org/apache/spark/sql/execution/python/InProcessArrowEvalPythonEvaluatorFactory.scala: ########## @@ -0,0 +1,551 @@ +/* + * 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.python + +import java.io.File +import java.nio.file.Files +import java.util.UUID +import java.util.concurrent.TimeUnit +import java.util.concurrent.atomic.AtomicInteger +import java.util.concurrent.locks.ReentrantLock + +import scala.collection.mutable.ArrayBuffer +import scala.jdk.CollectionConverters._ + +import com.google.common.util.concurrent.Uninterruptibles +import org.apache.arrow.c.{ArrowArray, ArrowSchema, BaseStruct} +import org.apache.arrow.util.AutoCloseables +import org.apache.arrow.vector.VectorSchemaRoot + +import org.apache.spark.{SparkEnv, SparkException, TaskContext} +import org.apache.spark.api.python.ChainedPythonFunctions +import org.apache.spark.memory.MemoryConsumer +import org.apache.spark.sql.catalyst.InternalRow +import org.apache.spark.sql.catalyst.expressions.{Attribute, Expression, JoinedRow, PythonUDF, UnsafeProjection, UnsafeRow} +import org.apache.spark.sql.execution.arrow.ArrowWriter +import org.apache.spark.sql.execution.metric.SQLMetric +import org.apache.spark.sql.execution.python.EvalPythonExec.ArgumentMetadata +import org.apache.spark.sql.types._ +import org.apache.spark.sql.util.ArrowUtils +import org.apache.spark.sql.vectorized.{ArrowColumnVector, ColumnarBatch, ColumnVector} +import org.apache.spark.util.Utils + +/** + * Evaluates scalar Python UDFs using Arrow CDI in the executor process. Only UDF arguments + * are converted to Arrow. Original rows are buffered in a spillable queue and joined with + * the results, unless all of them are UDF arguments that read back from Arrow unchanged. + * Each batch owns its Arrow buffers so Python can safely retain input arrays. + * + * The evaluator owns its queue, so that cleanup at task completion is coordinated with a + * consumer on another thread, such as a pipelined Python writer or a TRANSFORM feed thread. + */ +class InProcessArrowEvalPythonEvaluatorFactory( + childOutput: Seq[Attribute], + udfs: Seq[PythonUDF], + output: Seq[Attribute], + batchSize: Int, + maxBytes: Long, + timeZoneId: String, + largeVarTypes: Boolean, + hideTraceback: Boolean, + simplifiedTraceback: Boolean, + tracebackWithLocals: Boolean, + fullValidation: Boolean, + metrics: Map[String, SQLMetric]) + extends EvalPythonEvaluatorFactory(childOutput, udfs, output) { + + private[python] def runtimeSession: InProcessPythonRuntime.InterpreterSession = + InProcessPythonRuntime.currentSession + + /** Unused: `evaluateJoined` always evaluates the UDFs. */ + override protected def evaluate( + funcs: Seq[(ChainedPythonFunctions, Long)], + argMetas: Array[Array[ArgumentMetadata]], + rows: Iterator[InternalRow], + inputSchema: StructType, + context: TaskContext): Iterator[InternalRow] = + throw SparkException.internalError("In-process UDFs are evaluated with their input rows") + + override protected def evaluateJoined( + funcs: Seq[(ChainedPythonFunctions, Long)], + argMetas: Array[Array[ArgumentMetadata]], + rows: Iterator[InternalRow], + inputs: Seq[Expression], + inputSchema: StructType, + context: TaskContext): Option[Iterator[InternalRow]] = { + import InProcessArrowEvalPythonEvaluatorFactory.{Buffered, ReadBack, readsBack} + val inputColumns = inputs.length == childOutput.length && inputs.zip(childOutput).forall { + case (a: Attribute, c) => a.exprId == c.exprId + case _ => false + } + // If all input columns are UDF arguments, they are written to Arrow regardless. Read them + // back from the exported input vectors instead of buffering every input row, if their + // values read back from Arrow exactly as written and as fast as an unsafe row copy. + val joinInput = if (inputColumns && inputSchema.forall(f => readsBack(f.dataType))) { + ReadBack + } else if (inputColumns) { + Buffered(None) + } else { + // Each projected row is written to Arrow before the next input row is pulled, so the + // arguments go into a reused buffer rather than being copied value by value. + val projection = UnsafeProjection.create(inputs, childOutput) + projection.initialize(context.partitionId()) + Buffered(Some(projection)) + } + Some(evaluateBatches(funcs, argMetas, rows, inputSchema, context, joinInput)) + } + + private[python] def evaluateBatches( + funcs: Seq[(ChainedPythonFunctions, Long)], + argMetas: Array[Array[ArgumentMetadata]], + rows: Iterator[InternalRow], + inputSchema: StructType, + context: TaskContext, + joinInput: InProcessArrowEvalPythonEvaluatorFactory.JoinInput): Iterator[InternalRow] = { + import InProcessArrowEvalPythonEvaluatorFactory.{Buffered, ReadBack} + ArrowUtils.failDuplicatedFieldNames(inputSchema) + val functions = funcs.map { case (chain, _) => + if (chain.funcs.size != 1) { + throw SparkException.internalError( + "In-process UDF chains must use separate evaluation nodes") + } + chain.funcs.head + } + val inputOrdinals = argMetas.map(_.map(_.offset)) + def checkCancellation(): Unit = context.killTaskIfInterrupted() + + val expectedFields = udfs.map { udf => + ArrowUtils.toArrowField("result", udf.dataType, true, timeZoneId, largeVarTypes) + } + val processingTime = new InProcessArrowEvalPythonEvaluatorFactory.NanosecondTimer( + metrics("pythonProcessingTime")) + val initTime = new InProcessArrowEvalPythonEvaluatorFactory.NanosecondTimer( + metrics("pythonInitTime")) + val arrowSchema = ArrowUtils.toArrowSchema(inputSchema, timeZoneId, largeVarTypes) + // Capture before consuming input: an old task must never join a later context's session. + val runtime = runtimeSession + // Rows are copied out of the queue and Arrow vectors before they are returned, so they + // remain valid after task completion releases those, on whichever thread consumes them. + val resultProj = UnsafeProjection.create(output, output) + // Spill files go into a directory of the queue's own, created with the first disk queue, so + // that task completion can delete them when it cannot close the queue. + @volatile var spillDir: File = null + // Guarded by the queue's monitor. + var queueAbandoned = false + val (queue, projection) = joinInput match { + case Buffered(projection) => + val localDir = new File(Utils.getLocalDir(SparkEnv.get.conf)) + val serializerManager = SparkEnv.get.serializerManager + // Only the consumer holding the iterator's lock adds and removes rows. + val queue = new HybridRowQueue(context.taskMemoryManager(), localDir, + childOutput.length, serializerManager, lockFree = true) { + override protected def createDiskQueue(): RowQueue = synchronized { + if (spillDir == null) { + spillDir = Files.createTempDirectory(localDir.toPath, "inprocess-udf-").toFile + } + DiskRowQueue(Files.createTempFile(spillDir.toPath, "buffer", "").toFile, + childOutput.length, serializerManager) + } + + // Once task completion leaves the queue to the executor, it must not spill for other + // consumers into a directory that nothing deletes. + override def spill(size: Long, trigger: MemoryConsumer): Long = synchronized { + if (queueAbandoned) 0L else super.spill(size, trigger) + } + + // Queues of a task are distinct memory consumers, whatever their case-class fields. + override def equals(other: Any): Boolean = this eq other.asInstanceOf[AnyRef] + override def hashCode(): Int = System.identityHashCode(this) + override def canEqual(other: Any): Boolean = false + } + (queue, projection.orNull) + case ReadBack => (null, null) + } + val joined = new JoinedRow + val handles = functions.map(_ => UUID.randomUUID().toString) + var registered = false + var writer: ArrowWriter = null + val results = ArrayBuffer.empty[ArrowColumnVector] + var startedAt = 0L + + def closeBatch(): Unit = { + val resources = ArrayBuffer.empty[AutoCloseable] + resources ++= results + results.clear() + if (writer != null) { + resources += writer.root + writer = null + } + AutoCloseables.close(resources.asJava) + } + + val resources = new InProcessArrowEvalPythonEvaluatorFactory.IteratorResources( + hasTaskMemory = queue != null, + // Closing the queue deletes the spill files it tracks; deleteQuietly also removes any + // other, without starting a process or throwing, also on an interrupted thread. + releaseTaskMemory = () => if (queue != null) { + Utils.tryWithSafeFinally(queue.close())(Utils.deleteQuietly(spillDir)) + }, + abandonTaskMemory = () => if (queue != null) { + queue.synchronized { + queueAbandoned = true + Utils.deleteQuietly(spillDir) + } + }, + releaseOthers = () => { + if (startedAt != 0L) { + metrics("pythonTotalTime") += (System.nanoTime() - startedAt) / 1000000 + } + Utils.tryWithSafeFinally { + closeBatch() + } { + if (registered) runtime.release(handles) + } + }) + + context.addTaskCompletionListener[Unit](_ => resources.close()) + + new Iterator[InternalRow] { + private var batchIter: Iterator[InternalRow] = Iterator.empty + + private def endOfInput: Nothing = + throw new NoSuchElementException("End of in-process UDF input") + + // Releases the resources on failure without replacing its exception. + private def fail(t: Throwable): Nothing = + Utils.tryWithSafeFinally { throw t } { resources.close() } + + // Called with the lock held. + private def hasNextLocked: Boolean = { + if (startedAt == 0L) startedAt = System.nanoTime() + checkCancellation() + val available = batchIter.hasNext || { + resources.startReadingInput() + try !resources.isClosed && rows.hasNext finally resources.endReadingInput() + } + if (!available) resources.close() + available + } + + // Each call takes the lock without allocating a closure per row. Within a batch, only + // the consumer advances `batchIter`, whose `hasNext` compares row indexes. + override def hasNext: Boolean = { + if (batchIter.hasNext && !resources.isClosed) return true + if (!resources.enter()) return false + val available = try { + try hasNextLocked catch { case t: Throwable => fail(t) } + } finally { + resources.exit() + } + // Task completion may have closed the iterator while the input was being read. + available && !resources.isClosed + } + + override def next(): InternalRow = { + if (!resources.enter()) endOfInput + try { + try { + if (!hasNextLocked) endOfInput + if (!batchIter.hasNext) nextBatch() + val result = batchIter.next() + resultProj(if (queue != null) joined(queue.remove(), result) else result) + } catch { + case t: Throwable => fail(t) + } + } finally { + resources.exit() + } + } + + // Runs Python without the lock unless task completion already happened, and ends the + // input instead of returning the result if it happens meanwhile. + private def python[T](body: => T): T = { + if (resources.isClosed) endOfInput + val result = resources.withoutLock(body) + if (resources.isClosed) endOfInput + result + } + + /** + * Writes the next input row to the batch, returning false at the end of input or once + * task completion happened. If it happens while the row is read, the row is dropped. + */ + private def pullRow(): Boolean = { + resources.startReadingInput() + val row = try { + // Checked after marking, and again after `hasNext`, which may wait for input. + if (!resources.isClosed && rows.hasNext && !resources.isClosed) rows.next() else null Review Comment: Done in 3b69cd9: the fill loop no longer checks `isClosed`, since `startReadingInput()` does right after. A pulled row now takes three volatile reads: after marking, after `hasNext`, so that a close while waiting for input does not read another row, and after the read. The comment says that. -- 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]
