dtenedor commented on code in PR #44678:
URL: https://github.com/apache/spark/pull/44678#discussion_r1454199865
##########
sql/core/src/main/scala/org/apache/spark/sql/execution/python/EvalPythonUDTFExec.scala:
##########
@@ -137,4 +150,46 @@ trait EvalPythonUDTFExec extends UnaryExecNode {
}
}
}
+
+ lazy val memoryConsumer: Option[PythonUDTFMemoryConsumer] = {
+ if (TaskContext.get() != null) {
+ Some(PythonUDTFMemoryConsumer(udtf))
+ } else {
+ None
+ }
+ }
+}
+
+/**
+ * This class takes responsibility to allocate execution memory for UDTF
evaluation before it begins
+ * and free the memory after the evaluation is over.
+ *
+ * Background: If the UDTF's 'analyze' method returns an 'AnalyzeResult' with
a non-empty
+ * 'acquireExecutionMemoryMb' value, this value represents the amount of
memory in megabytes that
+ * the UDTF should request from each Spark executor that it runs on. Then the
UDTF takes
+ * responsibility to use at most this much memory, including all allocated
objects. The purpose of
+ * this functionality is to prevent executors from crashing by running out of
memory due to the
+ * extra memory consumption invoked by the UDTF's 'eval' and 'terminate' and
'cleanup' methods.
+ *
+ * In this class, Spark calls 'TaskMemoryManager.acquireExecutionMemory' with
the requested number
+ * of megabytes, and when Spark calls __init__ of the UDTF later, it updates
the
+ * acquiredExecutionMemory integer passed into the UDTF constructor to the
actual number returned
+ * from 'TaskMemoryManager.acquireExecutionMemory', so the 'eval' and
'terminate' and 'cleanup'
+ * methods know it and can ensure to bound memory usage to at most this number.
+ */
+case class PythonUDTFMemoryConsumer(udtf: PythonUDTF)
+ extends MemoryConsumer(TaskContext.get().taskMemoryManager(),
MemoryMode.ON_HEAP) {
+ private val BYTES_PER_MEGABYTE = 1024 * 1024
Review Comment:
Done.
##########
python/pyspark/sql/udtf.py:
##########
@@ -133,12 +133,28 @@ class AnalyzeResult:
If non-empty, this is a sequence of expressions that the UDTF is
specifying for Catalyst to
sort the input TABLE argument by. Note that the 'partitionBy' list
must also be non-empty
in this case.
+ acquireExecutionMemoryMbRequested: long
+ If this is not None, this represents the amount of memory in megabytes
that the UDTF should
+ request from each Spark executor that it runs on. Then the UDTF takes
responsibility to use
+ at most this much memory, including all allocated objects. The purpose
of this functionality
+ is to prevent executors from crashing by running out of memory due to
the extra memory
+ consumption invoked by the UDTF's 'eval' and 'terminate' and 'cleanup'
methods. Spark will
+ then call 'TaskMemoryManager.acquireExecutionMemory' with the
requested number of megabytes.
+ acquireExecutionMemoryMbActual: long
+ If there is a task context available, Spark will assign this field to
the number of
+ megabytes returned from the call to the
TaskMemoryManager.acquireExecutionMemory' method, as
+ consumed by the UDTF's'__init__' method. Therefore, its 'eval' and
'terminate' and 'cleanup'
+ methods will know it thereafter and can ensure to bound memory usage
to at most this number.
+ Note that there is no effect if the UDTF's 'analyze' method assigns a
value to this; it will
+ be overwritten.
"""
Review Comment:
Sounds good, I added a default of 100MB, which any UDTF can (and probably
should) override to a more specific number.
--
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]
