fix.

Author: gatorsmile

sql/core/src/main/scala/org/apache/spark/sql/execution/LocalTableScanExec.scala

@@ -28,12 +28,12 @@ import org.apache.spark.sql.execution.metric.SQLMetrics
  */
 case class LocalTableScanExec(
     output: Seq[Attribute],
-    rows: Seq[InternalRow]) extends LeafExecNode {
+    @transient rows: Seq[InternalRow]) extends LeafExecNode {
 
   override lazy val metrics = Map(
     "numOutputRows" -> SQLMetrics.createMetric(sparkContext, "number of output rows"))
 
-  private lazy val unsafeRows: Array[InternalRow] = {
+  @transient private lazy val unsafeRows: Array[InternalRow] = {
     if (rows.isEmpty) {
       Array.empty
     } else {

sql/core/src/test/scala/org/apache/spark/sql/execution/OptimizeMetadataOnlyQuerySuite.scala

@@ -117,4 +117,12 @@ class OptimizeMetadataOnlyQuerySuite extends QueryTest with SharedSQLContext {
     "select partcol1, max(partcol2) from srcpart where partcol1 = 0 group by rollup (partcol1)",
     "select partcol2 from (select partcol2 from srcpart where partcol1 = 0 union all " +
       "select partcol2 from srcpart where partcol1 = 1) t group by partcol2")
+
+  test("SPARK-21884 Fix StackOverflowError on MetadataOnlyQuery") {
+    withTable("t_1000") {
+      sql("CREATE TABLE t_1000 (a INT, p INT) USING PARQUET PARTITIONED BY (p)")
+      (1 to 1000).foreach(p => sql(s"ALTER TABLE t_1000 ADD PARTITION (p=$p)"))
+      sql("SELECT COUNT(DISTINCT p) FROM t_1000").collect()
+    }
+  }
 }