Use ContextAwareIterator to stop consuming after the task ends.

python/pyspark/sql/tests/test_pandas_map.py

@@ -15,9 +15,12 @@
 # limitations under the License.
 #
 import os
+import shutil
+import tempfile
 import time
 import unittest
 
+from pyspark.sql import Row
 from pyspark.testing.sqlutils import ReusedSQLTestCase, have_pandas, have_pyarrow, \
     pandas_requirement_message, pyarrow_requirement_message
 
@@ -112,6 +115,25 @@ def func(iterator):
         expected = df.collect()
         self.assertEquals(actual, expected)
 
+    # SPARK-33277
+    def test_map_in_pandas_with_column_vector(self):
+        path = tempfile.mkdtemp()
+        shutil.rmtree(path)
+
+        try:
+            self.spark.range(0, 200000, 1, 1).write.parquet(path)
+
+            def func(iterator):
+                for pdf in iterator:
+                    yield pd.DataFrame({'id': [0] * len(pdf)})
+
+            for offheap in ["true", "false"]:
+                with self.sql_conf({"spark.sql.columnVector.offheap.enabled": offheap}):
+                    self.assertEquals(
+                        self.spark.read.parquet(path).mapInPandas(func, 'id long').head(), Row(0))
+        finally:
+            shutil.rmtree(path)
+
 
 if __name__ == "__main__":
     from pyspark.sql.tests.test_pandas_map import *  # noqa: F401

python/pyspark/sql/tests/test_pandas_udf_scalar.py

@@ -1137,6 +1137,25 @@ def test_datasource_with_udf(self):
         finally:
             shutil.rmtree(path)
 
+    # SPARK-33277
+    def test_pandas_udf_with_column_vector(self):
+        path = tempfile.mkdtemp()
+        shutil.rmtree(path)
+
+        try:
+            self.spark.range(0, 200000, 1, 1).write.parquet(path)
+
+            @pandas_udf(LongType())
+            def udf(x):
+                return pd.Series([0] * len(x))
+
+            for offheap in ["true", "false"]:
+                with self.sql_conf({"spark.sql.columnVector.offheap.enabled": offheap}):
+                    self.assertEquals(
+                        self.spark.read.parquet(path).select(udf('id')).head(), Row(0))
+        finally:
+            shutil.rmtree(path)
+
 
 if __name__ == "__main__":
     from pyspark.sql.tests.test_pandas_udf_scalar import *  # noqa: F401

python/pyspark/sql/tests/test_udf.py

@@ -674,6 +674,26 @@ def test_udf_cache(self):
         self.assertEqual(df.select(udf(func)("id"))._jdf.queryExecution()
                          .withCachedData().getClass().getSimpleName(), 'InMemoryRelation')
 
+    # SPARK-33277
+    def test_udf_with_column_vector(self):
+        path = tempfile.mkdtemp()
+        shutil.rmtree(path)
+
+        try:
+            self.spark.range(0, 100000, 1, 1).write.parquet(path)
+
+            def f(x):
+                return 0
+
+            fUdf = udf(f, LongType())
+
+            for offheap in ["true", "false"]:
+                with self.sql_conf({"spark.sql.columnVector.offheap.enabled": offheap}):
+                    self.assertEquals(
+                        self.spark.read.parquet(path).select(fUdf('id')).head(), Row(0))
+        finally:
+            shutil.rmtree(path)
+
 
 class UDFInitializationTests(unittest.TestCase):
     def tearDown(self):

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

@@ -89,6 +89,7 @@ trait EvalPythonExec extends UnaryExecNode {
 
     inputRDD.mapPartitions { iter =>
       val context = TaskContext.get()
+      val contextAwareIterator = new ContextAwareIterator(iter, context)
 
       // The queue used to buffer input rows so we can drain it to
       // combine input with output from Python.
@@ -120,7 +121,7 @@ trait EvalPythonExec extends UnaryExecNode {
       }.toSeq)
 
       // Add rows to queue to join later with the result.
-      val projectedRowIter = iter.map { inputRow =>
+      val projectedRowIter = contextAwareIterator.map { inputRow =>
         queue.add(inputRow.asInstanceOf[UnsafeRow])
         projection(inputRow)
       }
@@ -137,3 +138,19 @@ trait EvalPythonExec extends UnaryExecNode {
     }
   }
 }
+
+/**
+ * A TaskContext aware iterator.
+ *
+ * As the Python evaluation consumes the parent iterator in a separate thread,
+ * it could consume more data from the parent even after the task ends and the parent is closed.
+ * Thus, we should use ContextAwareIterator to stop consuming after the task ends.
+ */
+private[spark] class ContextAwareIterator[IN](iter: Iterator[IN], context: TaskContext)
+  extends Iterator[IN] {
+
+  override def hasNext: Boolean =
+    !context.isCompleted() && !context.isInterrupted() && iter.hasNext
+
+  override def next(): IN = iter.next()
+}

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

@@ -61,16 +61,17 @@ case class MapInPandasExec(
       val pythonRunnerConf = ArrowUtils.getPythonRunnerConfMap(conf)
       val outputTypes = child.schema
 
+      val context = TaskContext.get()
+      val contextAwareIterator = new ContextAwareIterator(inputIter, context)
+
       // Here we wrap it via another row so that Python sides understand it
       // as a DataFrame.
-      val wrappedIter = inputIter.map(InternalRow(_))
+      val wrappedIter = contextAwareIterator.map(InternalRow(_))
 
       // DO NOT use iter.grouped(). See BatchIterator.
       val batchIter =
         if (batchSize > 0) new BatchIterator(wrappedIter, batchSize) else Iterator(wrappedIter)
 
-      val context = TaskContext.get()
-
       val columnarBatchIter = new ArrowPythonRunner(
         chainedFunc,
         PythonEvalType.SQL_MAP_PANDAS_ITER_UDF,