Fix inset for large queries. Disable record level filtering (apache-spark-on-k8s#74)

Author: pwoody

pom.xml

@@ -134,7 +134,7 @@
     <!-- Version used for internal directory structure -->
     <hive.version.short>1.2.1</hive.version.short>
     <derby.version>10.12.1.1</derby.version>
-    <parquet.version>1.9.0-palantir3</parquet.version>
+    <parquet.version>1.9.0-palantir4</parquet.version>
     <jetty.version>9.2.16.v20160414</jetty.version>
     <javaxservlet.version>3.1.0</javaxservlet.version>
     <chill.version>0.8.0</chill.version>

sql/core/src/main/java/org/apache/spark/sql/execution/datasources/parquet/SpecificParquetRecordReaderBase.java

@@ -22,6 +22,7 @@
 import static org.apache.parquet.format.converter.ParquetMetadataConverter.NO_FILTER;
 import static org.apache.parquet.format.converter.ParquetMetadataConverter.range;
 import static org.apache.parquet.hadoop.ParquetFileReader.readFooter;
+import static org.apache.parquet.hadoop.ParquetInputFormat.DICTIONARY_FILTERING_ENABLED;
 import static org.apache.parquet.hadoop.ParquetInputFormat.getFilter;
 
 import com.google.common.collect.ImmutableList;
@@ -107,8 +108,14 @@ public void initialize(InputSplit inputSplit, TaskAttemptContext taskAttemptCont
       footer = readFooter(configuration, file, range(split.getStart(), split.getEnd()));
       FilterCompat.Filter filter = getFilter(configuration);
       this.reader = ParquetFileReader.open(configuration, file, footer);
+      List<RowGroupFilter.FilterLevel> filterLevels =
+              ImmutableList.of(RowGroupFilter.FilterLevel.STATISTICS);
+      if (configuration.getBoolean(DICTIONARY_FILTERING_ENABLED, false)) {
+        filterLevels = ImmutableList.of(RowGroupFilter.FilterLevel.STATISTICS,
+                RowGroupFilter.FilterLevel.DICTIONARY);
+      }
       blocks = filterRowGroups(
-              ImmutableList.of(RowGroupFilter.FilterLevel.STATISTICS, RowGroupFilter.FilterLevel.DICTIONARY),
+              filterLevels,
               filter,
               footer.getBlocks(),
               reader);

sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/parquet/ParquetFileFormat.scala

@@ -394,6 +394,11 @@ class ParquetFileFormat
       SQLConf.PARQUET_INT96_AS_TIMESTAMP.key,
       int96AsTimestamp)
 
+    // By default, disable record level filtering.
+    if (hadoopConf.get(ParquetInputFormat.RECORD_FILTERING_ENABLED) == null) {
+      hadoopConf.setBoolean(ParquetInputFormat.RECORD_FILTERING_ENABLED, false)
+    }
+
     // Try to push down filters when filter push-down is enabled.
     val pushed =
       if (sparkSession.sessionState.conf.parquetFilterPushDown) {

sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/parquet/ParquetFilters.scala

@@ -30,6 +30,53 @@ import org.apache.spark.sql.types._
  */
 private[parquet] object ParquetFilters {
 
+  case class SetInFilter[T <: Comparable[T]](valueSet: Set[T])
+    extends UserDefinedPredicate[T] with Serializable {
+
+    override def keep(value: T): Boolean = {
+      value != null && valueSet.contains(value)
+    }
+
+    // Drop when no value in the set is within the statistics range.
+    override def canDrop(statistics: Statistics[T]): Boolean = {
+      val statMax = statistics.getMax
+      val statMin = statistics.getMin
+      val statRange = com.google.common.collect.Range.closed(statMin, statMax)
+      !valueSet.exists(value => statRange.contains(value))
+    }
+
+    // Can only drop not(in(set)) when we are know that every element in the block is in valueSet.
+    // From the statistics, we can only be assured of this when min == max.
+    override def inverseCanDrop(statistics: Statistics[T]): Boolean = {
+      val statMax = statistics.getMax
+      val statMin = statistics.getMin
+      statMin == statMax && valueSet.contains(statMin)
+    }
+  }
+
+  private val makeInSet: PartialFunction[DataType, (String, Set[Any]) => FilterPredicate] = {
+    case IntegerType =>
+      (n: String, v: Set[Any]) =>
+        FilterApi.userDefined(intColumn(n), SetInFilter(v.asInstanceOf[Set[java.lang.Integer]]))
+    case LongType =>
+      (n: String, v: Set[Any]) =>
+        FilterApi.userDefined(longColumn(n), SetInFilter(v.asInstanceOf[Set[java.lang.Long]]))
+    case FloatType =>
+      (n: String, v: Set[Any]) =>
+        FilterApi.userDefined(floatColumn(n), SetInFilter(v.asInstanceOf[Set[java.lang.Float]]))
+    case DoubleType =>
+      (n: String, v: Set[Any]) =>
+        FilterApi.userDefined(doubleColumn(n), SetInFilter(v.asInstanceOf[Set[java.lang.Double]]))
+    case StringType =>
+      (n: String, v: Set[Any]) =>
+        FilterApi.userDefined(binaryColumn(n),
+          SetInFilter(v.map(s => Binary.fromString(s.asInstanceOf[String]))))
+    case BinaryType =>
+      (n: String, v: Set[Any]) =>
+        FilterApi.userDefined(binaryColumn(n),
+          SetInFilter(v.map(e => Binary.fromReusedByteArray(e.asInstanceOf[Array[Byte]]))))
+  }
+
   private val makeEq: PartialFunction[DataType, (String, Any) => FilterPredicate] = {
     case BooleanType =>
       (n: String, v: Any) => FilterApi.eq(booleanColumn(n), v.asInstanceOf[java.lang.Boolean])
@@ -299,10 +346,7 @@ private[parquet] object ParquetFilters {
           .map(LogicalInverseRewriter.rewrite)
 
       case sources.In(name, values) if dataTypeOf.contains(name) =>
-        val eq = makeEq.lift(dataTypeOf(name))
-        values.flatMap { v =>
-          eq.map(_(name, v))
-        }.reduceLeftOption(FilterApi.or)
+        makeInSet.lift(dataTypeOf(name)).map(_(name, values.toSet))
 
       case _ => None
     }

sql/core/src/test/scala/org/apache/spark/sql/execution/datasources/parquet/ParquetFilterSuite.scala

@@ -24,6 +24,7 @@ import java.time.{LocalDate, ZoneId}
 import org.apache.parquet.filter2.predicate.{FilterPredicate, Operators}
 import org.apache.parquet.filter2.predicate.FilterApi._
 import org.apache.parquet.filter2.predicate.Operators.{Column => _, _}
+import org.apache.parquet.hadoop.{ParquetInputFormat, ParquetOutputFormat}
 
 import org.apache.spark.sql._
 import org.apache.spark.sql.catalyst.dsl.expressions._
@@ -111,7 +112,7 @@ class ParquetFilterSuite extends QueryTest with ParquetTest with SharedSQLContex
           // Doesn't bother checking type parameters here (e.g. `Eq[Integer]`)
           maybeFilter.exists(_.getClass === filterClass)
         }
-        checker(stripSparkFilter(query), expected)
+        checker(query, expected)
       }
     }
   }
@@ -557,11 +558,7 @@ class ParquetFilterSuite extends QueryTest with ParquetTest with SharedSQLContex
           (1 to 3).map(i => (i, i.toString)).toDF("a", "b").write.parquet(path)
           val df = spark.read.parquet(path).filter("a = 2")
 
-          // The result should be single row.
-          // When a filter is pushed to Parquet, Parquet can apply it to every row.
-          // So, we can check the number of rows returned from the Parquet
-          // to make sure our filter pushdown work.
-          assert(stripSparkFilter(df).count == 1)
+          assert(df.count == 1)
         }
       }
     }
@@ -676,20 +673,35 @@ class ParquetFilterSuite extends QueryTest with ParquetTest with SharedSQLContex
           val path = s"${dir.getCanonicalPath}/table1"
           (1 to 5).map(i => (i.toFloat, i%3)).toDF("a", "b").write.parquet(path)
           val df = spark.read.parquet(path).where("b in (0,2)")
-          assert(stripSparkFilter(df).count == 3)
+          assert(df.count == 3)
           val df1 = spark.read.parquet(path).where("not (b in (1))")
-          assert(stripSparkFilter(df1).count == 3)
+          assert(df1.count == 3)
           val df2 = spark.read.parquet(path).where("not (b in (1,3) or a <= 2)")
-          assert(stripSparkFilter(df2).count == 2)
+          assert(df2.count == 2)
           val df3 = spark.read.parquet(path).where("not (b in (1,3) and a <= 2)")
-          assert(stripSparkFilter(df3).count == 4)
+          assert(df3.count == 4)
           val df4 = spark.read.parquet(path).where("not (a <= 2)")
-          assert(stripSparkFilter(df4).count == 3)
+          assert(df4.count == 3)
         }
       }
     }
   }
 
+  test("Large In filters work with UDP") {
+    import testImplicits._
+    withSQLConf(ParquetOutputFormat.JOB_SUMMARY_LEVEL -> "ALL",
+      ParquetInputFormat.DICTIONARY_FILTERING_ENABLED -> "true") {
+      withTempPath { dir =>
+        val path = s"${dir.getCanonicalPath}/table1"
+        (1 to 1000).toDF().write.parquet(path)
+        val df = spark.read.parquet(path)
+        val filter = (1 to 499).map(i => i.toString).mkString(",")
+        assert(df.where(s"value in (${filter})").count() == 499)
+        assert(df.where(s"value not in (${filter})").count() == 501)
+      }
+    }
+  }
+
   test("Do not create Timestamp filters when interpreting from INT96") {
     val baseMillis = System.currentTimeMillis()
     def base(): Timestamp = new Timestamp(baseMillis)