dongjoon-hyun commented on code in PR #55552:
URL: https://github.com/apache/spark/pull/55552#discussion_r3155614673
##########
python/pyspark/worker.py:
##########
@@ -3588,12 +3588,93 @@ def process():
if hasattr(out_iter, "close"):
out_iter.close()
+ def pipelined_process():
+ """
+ Pipelined variant of process() that pre-fetches input batches in a
background
+ reader thread while the main thread computes the UDF and writes
output.
+ This allows input deserialization to overlap with UDF computation.
+ """
+ # Mark that pipelined mode is active so UDFs can verify the code
path.
+ os.environ["SPARK_PIPELINED_UDF_ACTIVE"] = "1"
+ import queue
+ import threading
+
+ queue_depth =
int(os.environ.get("SPARK_PIPELINED_UDF_QUEUE_DEPTH", "2"))
+ _SENTINEL = object()
+ input_queue = queue.Queue(maxsize=queue_depth)
+ reader_error = [None]
+ stop_event = threading.Event()
+
+ def _reader_thread():
+ try:
+ for batch in deserializer.load_stream(infile):
+ # Some serializers (e.g., ArrowStreamGroupSerializer,
+ # ArrowStreamAggPandasUDFSerializer) yield lazy
iterators
+ # that still read from infile. Materialize them here
so the
+ # main thread can consume them without touching infile.
+ if hasattr(batch, "__next__"):
+ batch = list(batch)
Review Comment:
Just a question. Any chance of OOM due to the memory peak due to this?
--
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]
