cloud-fan commented on code in PR #38468:
URL: https://github.com/apache/spark/pull/38468#discussion_r1019100196
##########
connector/connect/src/main/scala/org/apache/spark/sql/connect/service/SparkConnectStreamHandler.scala:
##########
@@ -114,10 +123,97 @@ 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 rows = dataframe.queryExecution.executedPlan.execute()
+ val numPartitions = rows.getNumPartitions
+ var numSent = 0
+
+ if (numPartitions > 0) {
+ type Batch = (Array[Byte], Long, Long)
+
+ val batches = rows.mapPartitionsInternal { iter =>
+ ArrowConverters
+ .toArrowBatchIterator(iter, schema, maxRecordsPerBatch, timeZoneId)
+ }
+
+ val signal = new Object
+ val partitions = Array.fill[Array[Batch]](numPartitions)(null)
+
+ val processPartition = (iter: Iterator[Batch]) => iter.toArray
+
+ val resultHandler = (partitionId: Int, partition: Array[Batch]) => {
+ signal.synchronized {
+ partitions(partitionId) = partition
+ signal.notify()
+ }
+ val i = 0 // Unit
+ }
+
+ spark.sparkContext.runJob(batches, processPartition, resultHandler)
+
+ var currentPartitionId = 0
+ while (currentPartitionId < numPartitions) {
+ val partition = signal.synchronized {
+ while (partitions(currentPartitionId) == null) {
+ signal.wait()
Review Comment:
The current approach still has the risk of holding all the results in the
driver memory (assuming the first partition comes last), which violates the
design goal of Spark Connect.
I think the Spark driver should send whichever partition that arrives to the
client, and the client should allocate an array to hold arrow batches of all
partitions. The client need to keep all the result in-memory anyway, so it's
better to ask the client to buffer the results ad reorder them by partition id.
--
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]
