[Spark] Skip non-deterministic filters in Data Skipping to prevent in…

…correct file pruning in Delta queries (#4141)

#### Which Delta project/connector is this regarding?
- [x] Spark
- [ ] Standalone
- [ ] Flink
- [ ] Kernel
- [ ] Other (fill in here)

## Description

This is a follow-up to the previous attempt to handle double-filtering
of non-deterministic conditions (e.g. rand() < 0.25) in
#4095.

It prevented non-deterministic filters from appearing in `unusedFilters`
in the `ScanReport` unless we added special pipelining for them from
`PrepareDeltaScan` to `filesForScan`.

This is also inconsistent with how we skip `subqueryFilters`. We now
treat `filesForScan` as the narrow waist to skip any filters.

## How was this patch tested?

UTs

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

No

spark/src/main/scala/org/apache/spark/sql/delta/stats/DataSkippingReader.scala

@@ -1222,14 +1222,16 @@ trait DataSkippingReaderBase
     import DeltaTableUtils._
     val partitionColumns = metadata.partitionColumns
 
-    // For data skipping, avoid using the filters that involve subqueries.
-
-    val (subqueryFilters, flatFilters) = filters.partition {
-      case f => containsSubquery(f)
+    // For data skipping, avoid using the filters that either:
+    // 1. involve subqueries.
+    // 2. are non-deterministic.
+    var (ineligibleFilters, eligibleFilters) = filters.partition {
+      case f => containsSubquery(f) || !f.deterministic
     }
 
-    val (partitionFilters, dataFilters) = flatFilters
-        .partition(isPredicatePartitionColumnsOnly(_, partitionColumns, spark))
+
+    val (partitionFilters, dataFilters) = eligibleFilters
+      .partition(isPredicatePartitionColumnsOnly(_, partitionColumns, spark))
 
     if (dataFilters.isEmpty) recordDeltaOperation(deltaLog, "delta.skipping.partition") {
       // When there are only partition filters we can scan allFiles
@@ -1246,7 +1248,7 @@ trait DataSkippingReaderBase
         dataFilters = ExpressionSet(Nil),
         partitionLikeDataFilters = ExpressionSet(Nil),
         rewrittenPartitionLikeDataFilters = Set.empty,
-        unusedFilters = ExpressionSet(subqueryFilters),
+        unusedFilters = ExpressionSet(ineligibleFilters),
         scanDurationMs = System.currentTimeMillis() - startTime,
         dataSkippingType =
           getCorrectDataSkippingType(DeltaDataSkippingType.partitionFilteringOnlyV1)
@@ -1323,7 +1325,7 @@ trait DataSkippingReaderBase
         dataFilters = ExpressionSet(skippingFilters.map(_._1)),
         partitionLikeDataFilters = ExpressionSet(partitionLikeFilters.map(_._1)),
         rewrittenPartitionLikeDataFilters = partitionLikeFilters.map(_._2.expr.expr).toSet,
-        unusedFilters = ExpressionSet(unusedFilters.map(_._1) ++ subqueryFilters),
+        unusedFilters = ExpressionSet(unusedFilters.map(_._1) ++ ineligibleFilters),
         scanDurationMs = System.currentTimeMillis() - startTime,
         dataSkippingType = getCorrectDataSkippingType(dataSkippingType)
       )

spark/src/main/scala/org/apache/spark/sql/delta/stats/PrepareDeltaScan.scala

@@ -114,22 +114,20 @@ trait PrepareDeltaScanBase extends Rule[LogicalPlan]
       limitOpt: Option[Int],
       filters: Seq[Expression],
       delta: LogicalRelation): DeltaScan = {
-    // Remove non-deterministic filters (e.g., rand() < 0.25) to prevent incorrect file pruning.
-    val deterministicFilters = filters.filter(_.deterministic)
     withStatusCode("DELTA", "Filtering files for query") {
       if (limitOpt.nonEmpty) {
         // If we trigger limit push down, the filters must be partition filters. Since
         // there are no data filters, we don't need to apply Generated Columns
         // optimization. See `DeltaTableScan` for more details.
-        return scanGenerator.filesForScan(limitOpt.get, deterministicFilters)
+        return scanGenerator.filesForScan(limitOpt.get, filters)
       }
       val filtersForScan =
         if (!GeneratedColumn.partitionFilterOptimizationEnabled(spark)) {
-          deterministicFilters
+          filters
         } else {
           val generatedPartitionFilters = GeneratedColumn.generatePartitionFilters(
-            spark, scanGenerator.snapshotToScan, deterministicFilters, delta)
-          deterministicFilters ++ generatedPartitionFilters
+            spark, scanGenerator.snapshotToScan, filters, delta)
+          filters ++ generatedPartitionFilters
         }
       scanGenerator.filesForScan(filtersForScan)
     }

spark/src/test/scala/org/apache/spark/sql/delta/stats/DataSkippingDeltaTests.scala

@@ -36,6 +36,7 @@ import org.apache.spark.sql._
 import org.apache.spark.sql.catalyst.QueryPlanningTracker
 import org.apache.spark.sql.catalyst.TableIdentifier
 import org.apache.spark.sql.catalyst.expressions.{Expression, Literal, PredicateHelper}
+import org.apache.spark.sql.catalyst.plans.logical.Filter
 import org.apache.spark.sql.functions.{col, lit}
 import org.apache.spark.sql.internal.SQLConf
 import org.apache.spark.sql.test.SharedSparkSession
@@ -1929,6 +1930,71 @@ trait DataSkippingDeltaTestsBase extends DeltaExcludedBySparkVersionTestMixinShi
     }
   }
 
+  test("File skipping with non-deterministic filters") {
+    withTable("tbl") {
+      // Create the table.
+      val df = spark.range(100).toDF()
+      df.write.mode("overwrite").format("delta").saveAsTable("tbl")
+
+      // Append 9 times to the table.
+      for (i <- 1 to 9) {
+        val df = spark.range(i * 100, (i + 1) * 100).toDF()
+        df.write.mode("append").format("delta").insertInto("tbl")
+      }
+
+      val query = "SELECT count(*) FROM tbl WHERE rand(0) < 0.25"
+      val result = sql(query).collect().head.getLong(0)
+      assert(result > 150, s"Expected around 250 rows (~0.25 * 1000), got: $result")
+
+      val predicates = sql(query).queryExecution.optimizedPlan.collect {
+        case Filter(condition, _) => condition
+      }.flatMap(splitConjunctivePredicates)
+      val scanResult = DeltaLog.forTable(spark, TableIdentifier("tbl"))
+        .update().filesForScan(predicates)
+      assert(scanResult.unusedFilters.nonEmpty)
+    }
+  }
+
+  test("File skipping with non-deterministic filters on partitioned tables") {
+    withTable("tbl_partitioned") {
+      import org.apache.spark.sql.functions.col
+
+      // Create initial DataFrame and add a partition column.
+      val df = spark.range(100).toDF().withColumn("p", col("id") % 10)
+      df.write
+        .mode("overwrite")
+        .format("delta")
+        .partitionBy("p")
+        .saveAsTable("tbl_partitioned")
+
+      // Append 9 more times to the table.
+      for (i <- 1 to 9) {
+        val newDF = spark.range(i * 100, (i + 1) * 100).toDF().withColumn("p", col("id") % 10)
+        newDF.write.mode("append").format("delta").insertInto("tbl_partitioned")
+      }
+
+      // Run query with a nondeterministic filter.
+      val query = "SELECT count(*) FROM tbl_partitioned WHERE rand(0) < 0.25"
+      val result = sql(query).collect().head.getLong(0)
+      // Assert that the row count is as expected (e.g., roughly 25% of rows).
+      assert(result > 150, s"Expected a reasonable number of rows, got: $result")
+
+      val predicates = sql(query).queryExecution.optimizedPlan.collect {
+        case Filter(condition, _) => condition
+      }.flatMap(splitConjunctivePredicates)
+      val scanResult = DeltaLog.forTable(spark, TableIdentifier("tbl_partitioned"))
+        .update().filesForScan(predicates)
+      assert(scanResult.unusedFilters.nonEmpty)
+
+      // Assert that entries are fetched from all 10 partitions
+      val distinctPartitions =
+        sql("SELECT DISTINCT p FROM tbl_partitioned WHERE rand(0) < 0.25")
+        .collect()
+        .length
+      assert(distinctPartitions == 10)
+    }
+  }
+
   protected def parse(deltaLog: DeltaLog, predicate: String): Seq[Expression] = {
 
     // We produce a wrong filter in this case otherwise