pan3793 commented on code in PR #38468:
URL: https://github.com/apache/spark/pull/38468#discussion_r1019048009
##########
connector/connect/src/main/scala/org/apache/spark/sql/connect/service/SparkConnectStreamHandler.scala:
##########
@@ -117,10 +129,91 @@ class SparkConnectStreamHandler(responseObserver:
StreamObserver[Response]) exte
responseObserver.onNext(response.build())
}
- responseObserver.onNext(sendMetricsToResponse(clientId, rows))
+ responseObserver.onNext(sendMetricsToResponse(clientId, dataframe))
responseObserver.onCompleted()
}
+ def processRowsAsArrowBatches(clientId: String, dataframe: DataFrame): Unit
= {
+ val spark = dataframe.sparkSession
+ val schema = dataframe.schema
+ // TODO: control the batch size instead of max records
+ val maxRecordsPerBatch = spark.sessionState.conf.arrowMaxRecordsPerBatch
+ val timeZoneId = spark.sessionState.conf.sessionLocalTimeZone
+
+ SQLExecution.withNewExecutionId(dataframe.queryExecution,
Some("collectArrow")) {
+ val pool =
ThreadUtils.newDaemonSingleThreadExecutor("connect-collect-arrow")
+ val tasks = collection.mutable.ArrayBuffer.empty[Future[_]]
+ val rows = dataframe.queryExecution.executedPlan.execute()
+
+ if (rows.getNumPartitions > 0) {
+ val batches = rows.mapPartitionsInternal { iter =>
+ ArrowConverters
+ .toArrowBatchIterator(iter, schema, maxRecordsPerBatch, timeZoneId)
+ }
+
+ val processPartition = (iter: Iterator[(Array[Byte], Long, Long)]) =>
iter.toArray
+
+ val resultHandler = (partitionId: Int, taskResult: Array[(Array[Byte],
Long, Long)]) => {
+ if (taskResult.exists(_._1.nonEmpty)) {
+ // only send non-empty partitions
+ val task = pool.submit(new Runnable {
+ override def run(): Unit = {
+ var batchId = partitionId.toLong << 33
+ taskResult.foreach { case (bytes, count, size) =>
+ val response =
proto.Response.newBuilder().setClientId(clientId)
+ val batch = proto.Response.ArrowBatch
+ .newBuilder()
+ .setBatchId(batchId)
+ .setRowCount(count)
+ .setUncompressedBytes(size)
+ .setCompressedBytes(bytes.length)
+ .setData(ByteString.copyFrom(bytes))
+ .build()
+ response.setArrowBatch(batch)
+ responseObserver.onNext(response.build())
Review Comment:
To be clear, my suggestion is to defer the deserialization and block
releasing to `JobWaiter#taskSucceeded` phase.
The second way you suggested has many performance benefits but requires the
client to communicate with the external storage service.
The `IndirectTaskResult` way leverages the existing code and spark build-in
block mechanism to transfer data, we can benefit w/ a little code modification,
and we don't need to worry about result cleanup.
--
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]
