[SPARK-37779][SQL] Make ColumnarToRowExec plan canonicalizable after …

…(de)serialization

This PR proposes to add a driver-side check on `supportsColumnar` sanity check at `ColumnarToRowExec`.

SPARK-23731 fixed the plans to be serializable by leveraging lazy but SPARK-28213 happened to refer to the lazy variable at: https://github.com/apache/spark/blob/77b164aac9764049a4820064421ef82ec0bc14fb/sql/core/src/main/scala/org/apache/spark/sql/execution/Columnar.scala#L68

This can fail during canonicalization during, for example, eliminating sub common expressions (on executor side):

```
java.lang.NullPointerException
    at org.apache.spark.sql.execution.FileSourceScanExec.supportsColumnar$lzycompute(DataSourceScanExec.scala:280)
    at org.apache.spark.sql.execution.FileSourceScanExec.supportsColumnar(DataSourceScanExec.scala:279)
    at org.apache.spark.sql.execution.InputAdapter.supportsColumnar(WholeStageCodegenExec.scala:509)
    at org.apache.spark.sql.execution.ColumnarToRowExec.<init>(Columnar.scala:67)
    ...
    at org.apache.spark.sql.catalyst.plans.QueryPlan.canonicalized$lzycompute(QueryPlan.scala:581)
    at org.apache.spark.sql.catalyst.plans.QueryPlan.canonicalized(QueryPlan.scala:580)
    at org.apache.spark.sql.execution.ScalarSubquery.canonicalized$lzycompute(subquery.scala:110)
    ...
    at org.apache.spark.sql.catalyst.expressions.ExpressionEquals.hashCode(EquivalentExpressions.scala:275)
    ...
    at scala.collection.mutable.HashTable.findEntry$(HashTable.scala:135)
    at scala.collection.mutable.HashMap.findEntry(HashMap.scala:44)
    at scala.collection.mutable.HashMap.get(HashMap.scala:74)
    at org.apache.spark.sql.catalyst.expressions.EquivalentExpressions.addExpr(EquivalentExpressions.scala:46)
    at org.apache.spark.sql.catalyst.expressions.EquivalentExpressions.addExprTreeHelper$1(EquivalentExpressions.scala:147)
    at org.apache.spark.sql.catalyst.expressions.EquivalentExpressions.addExprTree(EquivalentExpressions.scala:170)
    at org.apache.spark.sql.catalyst.expressions.SubExprEvaluationRuntime.$anonfun$proxyExpressions$1(SubExprEvaluationRuntime.scala:89)
    at org.apache.spark.sql.catalyst.expressions.SubExprEvaluationRuntime.$anonfun$proxyExpressions$1$adapted(SubExprEvaluationRuntime.scala:89)
    at scala.collection.immutable.List.foreach(List.scala:392)
```

This fix is still a bandaid fix but at least addresses the issue with minimized change - this fix should ideally be backported too.

Pretty unlikely - when `ColumnarToRowExec` has to be canonicalized on the executor side (see the stacktrace), but yes. it would fix a bug.

Unittest was added.

Closes #35058 from HyukjinKwon/SPARK-37779.

Authored-by: Hyukjin Kwon <gurwls223@apache.org>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
(cherry picked from commit 195f1aa)
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>

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

@@ -64,7 +64,8 @@ trait ColumnarToRowTransition extends UnaryExecNode
  * [[MapPartitionsInRWithArrowExec]]. Eventually this should replace those implementations.
  */
 case class ColumnarToRowExec(child: SparkPlan) extends ColumnarToRowTransition with CodegenSupport {
-  assert(child.supportsColumnar)
+  // supportsColumnar requires to be only called on driver side, see also SPARK-37779.
+  assert(TaskContext.get != null || child.supportsColumnar)
 
   override def output: Seq[Attribute] = child.output
 

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

@@ -88,4 +88,23 @@ class SparkPlanSuite extends QueryTest with SharedSparkSession {
   test("SPARK-30780 empty LocalTableScan should use RDD without partitions") {
     assert(LocalTableScanExec(Nil, Nil).execute().getNumPartitions == 0)
   }
+
+  test("SPARK-37779: ColumnarToRowExec should be canonicalizable after being (de)serialized") {
+    withSQLConf(SQLConf.USE_V1_SOURCE_LIST.key -> "parquet") {
+      withTempPath { path =>
+        spark.range(1).write.parquet(path.getAbsolutePath)
+        val df = spark.read.parquet(path.getAbsolutePath)
+        val columnarToRowExec =
+          df.queryExecution.executedPlan.collectFirst { case p: ColumnarToRowExec => p }.get
+        try {
+          spark.range(1).foreach { _ =>
+            columnarToRowExec.canonicalized
+            ()
+          }
+        } catch {
+          case e: Throwable => fail("ColumnarToRowExec was not canonicalizable", e)
+        }
+      }
+    }
+  }
 }