[SPARK-51655][SQL] Fix metric collection in UnionLoopExec and add test

### What changes were proposed in this pull request?

Fix metrics collection method for Recursive CTEs. Also, add new metric which tracks the number of rows that the anchor returns.

### Why are the changes needed?

Current way of collecting metrics for recursive CTEs is incorrect.

### Does this PR introduce _any_ user-facing change?

No.

### How was this patch tested?

New test in SQLMetricsSuite.

### Was this patch authored or co-authored using generative AI tooling?

No.

Closes #50449 from Pajaraja/pavle-martinovic_data/MetricFixAndTestRecursiveCTE.

Authored-by: pavle-martinovic_data <[email protected]>
Signed-off-by: Wenchen Fan <[email protected]>

Author: pavle-martinovic_data

Diff for: sql/core/src/main/scala/org/apache/spark/sql/execution/UnionLoopExec.scala

@@ -91,7 +91,8 @@ case class UnionLoopExec(
 
   override lazy val metrics = Map(
     "numOutputRows" -> SQLMetrics.createMetric(sparkContext, "number of output rows"),
-    "numIterations" -> SQLMetrics.createMetric(sparkContext, "number of recursive iterations"))
+    "numIterations" -> SQLMetrics.createMetric(sparkContext, "number of recursive iterations"),
+    "numAnchorOutputRows" -> SQLMetrics.createMetric(sparkContext, "number of anchor output rows"))
 
   /**
    * This function executes the plan (optionally with appended limit node) and caches the result,
@@ -123,6 +124,7 @@ case class UnionLoopExec(
     val executionId = sparkContext.getLocalProperty(SQLExecution.EXECUTION_ID_KEY)
     val numOutputRows = longMetric("numOutputRows")
     val numIterations = longMetric("numIterations")
+    val numAnchorOutputRows = longMetric("numAnchorOutputRows")
     val levelLimit = conf.getConf(SQLConf.CTE_RECURSION_LEVEL_LIMIT)
     val rowLimit = conf.getConf(SQLConf.CTE_RECURSION_ROW_LIMIT)
 
@@ -136,6 +138,8 @@ case class UnionLoopExec(
 
     var (prevDF, prevCount) = executeAndCacheAndCount(anchor, currentLimit)
 
+    numAnchorOutputRows += prevCount
+
     var currentLevel = 1
 
     var currentNumRows = 0
@@ -177,7 +181,6 @@ case class UnionLoopExec(
       // Update metrics
       numOutputRows += prevCount
       numIterations += 1
-      SQLMetrics.postDriverMetricUpdates(sparkContext, executionId, metrics.values.toSeq)
 
       if (!limitReached) {
         // the current plan is created by substituting UnionLoopRef node with the project node of
@@ -200,6 +203,8 @@ case class UnionLoopExec(
       }
     }
 
+    SQLMetrics.postDriverMetricUpdates(sparkContext, executionId, metrics.values.toSeq)
+
     if (unionChildren.isEmpty) {
       new EmptyRDD[InternalRow](sparkContext)
     } else {

Diff for: sql/core/src/test/scala/org/apache/spark/sql/execution/metric/SQLMetricsSuite.scala

@@ -107,6 +107,24 @@ class SQLMetricsSuite extends SharedSparkSession with SQLMetricsTestUtils
     }
   }
 
+  test("Recursive CTEs metrics") {
+    val df = sql("""WITH RECURSIVE t(n) AS(
+                          | VALUES 1, 2
+                          | UNION ALL
+                          | SELECT n+1 FROM t WHERE n < 20
+                          | )
+                          | SELECT * FROM t""".stripMargin)
+    val unionLoopExec = df.queryExecution.executedPlan.collect {
+      case ule: UnionLoopExec => ule
+    }
+    sparkContext.listenerBus.waitUntilEmpty()
+    assert(unionLoopExec.size == 1)
+    val expected = Map("number of output rows" -> 39L, "number of recursive iterations" -> 20L,
+      "number of anchor output rows" -> 2L)
+    testSparkPlanMetrics(df, 22, Map(
+      2L -> (("UnionLoop", expected))))
+  }
+
   test("Filter metrics") {
     // Assume the execution plan is
     // PhysicalRDD(nodeId = 1) -> Filter(nodeId = 0)