Remove the Arrow allocation cleanup routine in TaskCompletionListen to avoid race condition

Author: icexelloss

sql/core/src/main/scala/org/apache/spark/sql/execution/python/ArrowPythonRunner.scala

@@ -70,19 +70,13 @@ class ArrowPythonRunner(
         val arrowSchema = ArrowUtils.toArrowSchema(schema, timeZoneId)
         val allocator = ArrowUtils.rootAllocator.newChildAllocator(
           s"stdout writer for $pythonExec", 0, Long.MaxValue)
-
         val root = VectorSchemaRoot.create(arrowSchema, allocator)
-        val arrowWriter = ArrowWriter.create(root)
-
-        context.addTaskCompletionListener { _ =>
-          root.close()
-          allocator.close()
-        }
-
-        val writer = new ArrowStreamWriter(root, null, dataOut)
-        writer.start()
 
         Utils.tryWithSafeFinally {
+          val arrowWriter = ArrowWriter.create(root)
+          val writer = new ArrowStreamWriter(root, null, dataOut)
+          writer.start()
+
           while (inputIterator.hasNext) {
             val nextBatch = inputIterator.next()
 
@@ -94,8 +88,18 @@ class ArrowPythonRunner(
             writer.writeBatch()
             arrowWriter.reset()
           }
-        } {
           writer.end()
+        } {
+          // If we close root and allocator in TaskCompletionListener, there could be a race
+          // condition where the writer thread keeps writing to the VectorSchemaRoot while
+          // it's being closed by the TaskCompletion listener.
+          // Closing root and allocator here is cleaner because root and allocator is owned
+          // by the writer thread and is only visible to the writer thread.
+          //
+          // If the writer thread is interrupted by TaskCompletionListener, it should either
+          // (1) in the try block, in which case it will get an InterruptedException when
+          // performing io, and goes into the finally block or (2) in the finally block,
+          // in which case it will ignore the interruption and close the resources.
           root.close()
           allocator.close()
         }