sunchao commented on code in PR #6988: URL: https://github.com/apache/datafusion-comet/pull/6988#discussion_r4272822039
########## spark/src/main/scala/org/apache/spark/sql/comet/execution/arrow/CacheBatchCoalescer.scala: ########## @@ -0,0 +1,216 @@ +/* + * 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.comet.execution.arrow + +import scala.collection.JavaConverters._ +import scala.collection.mutable +import scala.util.control.NonFatal + +import org.apache.arrow.memory.BufferAllocator +import org.apache.arrow.vector.VectorSchemaRoot +import org.apache.spark.TaskContext +import org.apache.spark.sql.catalyst.expressions.Attribute +import org.apache.spark.sql.comet.util.Utils +import org.apache.spark.sql.vectorized.ColumnarBatch + +import org.apache.comet.vector.NativeUtil + +/** + * The columnar input of a cache build, with its small batches combined. + * + * The columnar write stores each batch it is handed as one cached batch, and nothing upstream + * bounds how small a batch can be. A shuffle read returns at least one batch per map output + * block, so a relation cached from a shuffle over many map tasks would otherwise hold thousands + * of batches of a few hundred rows. Each carries its own message metadata, compression framing + * and statistics row, and every read of the cache pays for the batch count. + * + * A batch the writer can store as it stands, Arrow-backed and laid out the way the reader reads + * it, is passed through when it holds at least half of `targetRows`, so full native batches are + * still written without a copy. Every other batch is copied into batches of `targetRows` rows, + * the size the row path cuts at, and split where it crosses that size. A small batch is worth the + * copy, and a batch the writer cannot store as it stands has to be copied into Arrow anyway, so + * combining it costs nothing more. The copy reads each column through the `ColumnVector` + * interface, as `ArrowWriter.writeColumns` does for any input, so it handles dictionary-encoded + * and nested columns, and offsets that do not start at zero, as a native producer's sliced arrays + * can have. Arrow's `VectorAppender` assumes they start at zero. + * + * Row order is kept: the rows copied so far are emitted before a batch that passes through. Each + * batch is copied before the input is advanced, because a native producer releases the batch it + * returned last when it is asked for the next one. + * + * The batches built here are owned here. Each is released when the next batch is requested, and + * whatever is still open when the task completes is released then. A batch passed through still + * belongs to its producer. + */ +private[arrow] class CacheBatchCoalescer( + input: Iterator[ColumnarBatch], + attrs: Seq[Attribute], + targetRows: Int, + allocator: BufferAllocator) + extends Iterator[ColumnarBatch] { + + require(targetRows > 0, s"The target batch size must be positive, but was $targetRows") + + private val arrowSchema = + Utils.toArrowSchema(Utils.fromAttributes(attrs), CometArrowStream.NATIVE_TIMEZONE) + private val readerFields = arrowSchema.getFields.asScala.toIndexedSeq + + // A batch at least this large that the writer can store as it stands is not copied. + private val passThroughRows = math.max(1, targetRows / 2) + + // The batch being filled. + private var pending: VectorSchemaRoot = null + private var pendingWriter: ArrowWriter = null + private var pendingRows = 0 + + // Batches built here and filled, in input order, ahead of `held`. + private val filled = mutable.Queue.empty[VectorSchemaRoot] + + // A batch that passes through, waiting for the rows copied ahead of it to be emitted. + private var held: ColumnarBatch = null + + // The batch the next call to `next` returns, and the root behind it when it was built here. + private var ready: ColumnarBatch = null + private var readyRoot: VectorSchemaRoot = null + + // The root behind the batch `next` returned last, released once the consumer asks for more. + private var emitted: VectorSchemaRoot = null + + private var closed = false + + Option(TaskContext.get()).foreach(_.addTaskCompletionListener[Unit](_ => close())) + + override def hasNext: Boolean = { + if (ready == null && !closed) { + release(emitted) + emitted = null + advance() + } + ready != null + } + + override def next(): ColumnarBatch = { + if (!hasNext) { + throw new NoSuchElementException("No more cache input batches") + } + val batch = ready + emitted = readyRoot + ready = null + readyRoot = null + batch + } + + /** Releases every batch built here that is still open. Idempotent. */ + def close(): Unit = { + if (!closed) { + closed = true + release(emitted) + release(readyRoot) + release(pending) + filled.foreach(release) + filled.clear() + emitted = null + ready = null + readyRoot = null + pending = null + pendingWriter = null + held = null + } + } + + private def advance(): Unit = { + while (filled.isEmpty && held == null && input.hasNext) { + val batch = input.next() + if (passesThrough(batch)) { + finishPending() + held = batch + } else { + append(batch) + } + } + if (filled.isEmpty && held == null) { + // The input is exhausted. + finishPending() + } + if (filled.nonEmpty) { + readyRoot = filled.dequeue() + ready = NativeUtil.rootAsBatch(readyRoot) + } else if (held != null) { + ready = held + held = null + } + } + + private def passesThrough(batch: ColumnarBatch): Boolean = + batch.numRows() >= passThroughRows && + Utils.isArrowBacked(batch) && + CachedBatchIpc.matchesReaderLayout(batch, readerFields) + + // Copies every row of `batch`, filling batches of targetRows rows. + private def append(batch: ColumnarBatch): Unit = { + val numRows = batch.numRows() + var start = 0 + while (start < numRows) { Review Comment: [P2] Emit split fragments incrementally instead of queuing the entire input. When a Spark-vector batch exceeds `targetRows`, this loop allocates and retains every output root before `advance()` returns the first one. Each variable-width vector receives Arrow's initial buffer allocation, so reducing `spark.sql.inMemoryColumnarStorage.batchSize` can greatly increase memory usage. For a normal-sized 4096-row input with four one-byte string columns, the previous conversion held 256 KiB, while a target of 100 retained 7.69 MiB and a target of 10 exhausted a bounded 64 MiB allocator before returning any output. Production uses the unbounded `CometArrowAllocator`, so this additional off-heap memory bypasses Spark's memory budget. Please retain the current input and row offset, emit one completed fragment, and continue copying only when the consumer requests another. The producer can remain unadvanced until that input is exhausted. Evidence: Reproduced with the exact-head `CacheBatchCoalescer` and `ArrowWriters`, Spark 4.1.3, Arrow 18.3.0, and `-Darrow.allocation.manager.type=Unsafe`. Construct four `OnHeapColumnVector(4096, StringType)` columns containing the single byte 'x', wrap them in a `ColumnarBatch`, and call `new CacheBatchCoalescer(Iterator(batch), attrs, targetRows, new RootAllocator(64L * 1024 * 1024)).next()`. The unchanged base conversion sequence (`ArrowWriter.create(root, 4096)`, `writeColumns(batch, 0, 4096)`, `finish()`) retained 262144 bytes. Head retained 8060928 bytes at target 100. At target 10 it threw `OutOfMemoryException` before the first output, with 67092480 bytes allocated. Explicit cleanup returned allocation to zero. Reproduction and logs are under `/tmp/pr6988-review/`. -- 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]
