Yicong-Huang commented on code in PR #55716:
URL: https://github.com/apache/spark/pull/55716#discussion_r3220425346
##########
python/pyspark/worker.py:
##########
@@ -3574,11 +3578,20 @@ def func(_, it):
return func, None, ser, ser
-@with_faulthandler
-def main(infile, outfile):
+def invoke_udf(message_receiver: SparkMessageReceiver, outfile: BinaryIO):
+ """
+ This function is the main processing function for worker.py.
+ It receives messages from the JVM, processes the data, and sends back
results.
+ This method goes through three phases:
+
+ Initialization -> Processing -> Finish/Cleanup
+ """
try:
boot_time = time.time()
+ # Initialization
+ infile = message_receiver.get_init_message()
Review Comment:
`infile` is misleading now. shall we change it to `init_message`?
##########
python/pyspark/worker.py:
##########
@@ -3602,6 +3615,10 @@ def main(infile, outfile):
runner_conf = RunnerConf(init_info.runner_conf)
eval_conf = EvalConf(init_info.eval_conf)
if eval_type == PythonEvalType.NON_UDF:
+ # The type checker needs some help here..
+ # See the code in WorkerInitInfo.from_stream(infile)
+ # to see the correct type.
+ assert isinstance(init_info.udf_info, memoryview)
Review Comment:
I don't think adding runtime assertions is a good idea. to make mypy happy,
maybe `typing.cast(memoryview, init_info.udf_info)`?
##########
python/pyspark/messages/socket/spark_socket_message_receiver.py:
##########
@@ -0,0 +1,64 @@
+#
+# 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.
+#
+
+from typing import BinaryIO
+
+from pyspark.serializers import read_int, SpecialLengths
+from pyspark.messages.zero_copy_byte_stream import ZeroCopyByteStream
+from pyspark.messages.spark_message_receiver import (
+ SparkMessageReceiver,
+)
+
+
+def _assert_message_id(message_id: int, expected: int) -> None:
+ assert message_id == expected, (
+ f"Expected message with id {expected} " + f"but got message with id
{message_id} instead."
+ )
+
+
+class SparkSocketMessageReceiver(SparkMessageReceiver):
+ def __init__(self, infile: BinaryIO):
+ super().__init__()
+ self._infile = infile
+
+ def _do_get_init_message(self) -> ZeroCopyByteStream:
+ message_id = read_int(self._infile)
+ _assert_message_id(message_id, SpecialLengths.START_OF_INIT_MESSAGE)
+
+ # Read the length and init content
+ message_length = read_int(self._infile)
+ message_content = self._infile.read(message_length)
+
+ return ZeroCopyByteStream(memoryview(message_content))
+
+ def _do_get_data_stream(self) -> BinaryIO:
+ # For socket communication, we just pass along the underlying socket
+ # for the data channel. We already stripped the initialization data
+ # at this state. Therefore, any bytes following this are data bytes.
+ #
+ # Note: We deliberately did not introduce a message header for
+ # data messages to reduce the overhead, especially for small
+ # batch sizes and real-time-mode (RTM).
+ return self._infile
+
+ def _do_is_stream_finished(self) -> bool:
+ # Check if the stream is finished.
+ # If everything finished properly, we should read a
+ # 'END_OF_STREAM'. If we read somethign else this means
Review Comment:
`something`
##########
python/pyspark/worker.py:
##########
@@ -3652,14 +3675,22 @@ def process():
send_accumulator_updates(outfile)
# check end of stream
- if read_int(infile) == SpecialLengths.END_OF_STREAM:
+ if message_receiver.is_stream_finished():
write_int(SpecialLengths.END_OF_STREAM, outfile)
else:
# write a different value to tell JVM to not reuse this worker
write_int(SpecialLengths.END_OF_DATA_SECTION, outfile)
sys.exit(-1)
+@with_faulthandler
+def main(infile, outfile):
+ # Instantiate socket message readers for executing the UDF
+ socket_reader = SparkSocketMessageReceiver(infile)
Review Comment:
I still prefer we have a flag and by default use the original code path.
replacing it directly would need some benchmark proof.
##########
python/pyspark/messages/spark_message_receiver.py:
##########
@@ -0,0 +1,126 @@
+#
+# 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.
+#
+
+from enum import Enum
+from functools import wraps
+from typing import BinaryIO, Callable, TypeVar
+from abc import ABC, abstractmethod
+
+from pyspark.messages.zero_copy_byte_stream import ZeroCopyByteStream
+
+
+T = TypeVar("T", bound="SparkMessageReceiver")
+R = TypeVar("R")
+
+
+class MessageState(Enum):
+ WAITING_FOR_INIT = 1
+ WAITING_FOR_DATA = 2
+ WAITING_FOR_FINISH = 3
+ DONE = 4
+
+
+class SparkMessageReceiver(ABC):
+ """
+ Generic class that implements receiving messages from Spark.
+ Caution: This class is STATEFUL. It is expected, that the
+ methods of this class are called in the following order:
Review Comment:
do we have tests for this receiver and its implementations?
As it contains some sort of state machine, it would be necessary to make
sure the state transition are enforced and verified. For instance, it should
not get two inits.
--
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]
