hvanhovell commented on code in PR #38468:
URL: https://github.com/apache/spark/pull/38468#discussion_r1019036963
##########
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:
You could do that yes. If we get the format right we could just serve the
block directly to the client (to avoid deserialization).
Alternatively you could also write the data to some persistent storage on
the executors, and just pass the file paths to the client. That will be the
most efficient from the POV of the driver. This requires you have some garbage
collection in place to clean-up results.
--
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]
