[SPARK-6910] [SQL] Support for pushing predicates down to metastore for partition pruning

This PR supersedes my old one apache#6921. Since my patch has changed quite a bit, I am opening a new PR to make it easier to review.

The changes include-
* Implement `toMetastoreFilter()` function in `HiveShim` that takes `Seq[Expression]` and converts them into a filter string for Hive metastore.
 * This functions matches all the `AttributeReference` + `BinaryComparisonOp` + `Integral/StringType` patterns in `Seq[Expression]` and fold them into a string.
* Change `hiveQlPartitions` field in `MetastoreRelation` to `getHiveQlPartitions()` function that takes a filter string parameter.
* Call `getHiveQlPartitions()` in `HiveTableScan` with a filter string.

But there are some cases in which predicate pushdown is disabled-

Case | Predicate pushdown
------- | -----------------------------
Hive integral and string types | Yes
Hive varchar type | No
Hive 0.13 and newer | Yes
Hive 0.12 and older | No
convertMetastoreParquet=false | Yes
convertMetastoreParquet=true | No

In case of `convertMetastoreParquet=true`, predicates are not pushed down because this conversion happens in an `Analyzer` rule (`HiveMetastoreCatalog.ParquetConversions`). At this point, `HiveTableScan` hasn't run, so predicates are not available. But reading the source code, I think it is intentional to convert the entire Hive table w/ all the partitions into `ParquetRelation` because then `ParquetRelation` can be cached and reused for any query against that table. Please correct me if I am wrong.

cc marmbrus

Author: Cheolsoo Park <[email protected]>

Closes apache#7216 from piaozhexiu/SPARK-6910-2 and squashes the following commits:

aa1490f [Cheolsoo Park] Fix ordering of imports
c212c4d [Cheolsoo Park] Incorporate review comments
5e93f9d [Cheolsoo Park] Predicate pushdown into Hive metastore

Author: Cheolsoo Park

sql/hive/src/main/scala/org/apache/spark/sql/hive/HiveMetastoreCatalog.scala

@@ -301,7 +301,9 @@ private[hive] class HiveMetastoreCatalog(val client: ClientInterface, hive: Hive
     val result = if (metastoreRelation.hiveQlTable.isPartitioned) {
       val partitionSchema = StructType.fromAttributes(metastoreRelation.partitionKeys)
       val partitionColumnDataTypes = partitionSchema.map(_.dataType)
-      val partitions = metastoreRelation.hiveQlPartitions.map { p =>
+      // We're converting the entire table into ParquetRelation, so predicates to Hive metastore
+      // are empty.
+      val partitions = metastoreRelation.getHiveQlPartitions().map { p =>
         val location = p.getLocation
         val values = InternalRow.fromSeq(p.getValues.zip(partitionColumnDataTypes).map {
           case (rawValue, dataType) => Cast(Literal(rawValue), dataType).eval(null)
@@ -644,32 +646,6 @@ private[hive] case class MetastoreRelation
     new Table(tTable)
   }
 
-  @transient val hiveQlPartitions: Seq[Partition] = table.getAllPartitions.map { p =>
-    val tPartition = new org.apache.hadoop.hive.metastore.api.Partition
-    tPartition.setDbName(databaseName)
-    tPartition.setTableName(tableName)
-    tPartition.setValues(p.values)
-
-    val sd = new org.apache.hadoop.hive.metastore.api.StorageDescriptor()
-    tPartition.setSd(sd)
-    sd.setCols(table.schema.map(c => new FieldSchema(c.name, c.hiveType, c.comment)))
-
-    sd.setLocation(p.storage.location)
-    sd.setInputFormat(p.storage.inputFormat)
-    sd.setOutputFormat(p.storage.outputFormat)
-
-    val serdeInfo = new org.apache.hadoop.hive.metastore.api.SerDeInfo
-    sd.setSerdeInfo(serdeInfo)
-    serdeInfo.setSerializationLib(p.storage.serde)
-
-    val serdeParameters = new java.util.HashMap[String, String]()
-    serdeInfo.setParameters(serdeParameters)
-    table.serdeProperties.foreach { case (k, v) => serdeParameters.put(k, v) }
-    p.storage.serdeProperties.foreach { case (k, v) => serdeParameters.put(k, v) }
-
-    new Partition(hiveQlTable, tPartition)
-  }
-
   @transient override lazy val statistics: Statistics = Statistics(
     sizeInBytes = {
       val totalSize = hiveQlTable.getParameters.get(StatsSetupConst.TOTAL_SIZE)
@@ -690,6 +666,34 @@ private[hive] case class MetastoreRelation
     }
   )
 
+  def getHiveQlPartitions(predicates: Seq[Expression] = Nil): Seq[Partition] = {
+    table.getPartitions(predicates).map { p =>
+      val tPartition = new org.apache.hadoop.hive.metastore.api.Partition
+      tPartition.setDbName(databaseName)
+      tPartition.setTableName(tableName)
+      tPartition.setValues(p.values)
+
+      val sd = new org.apache.hadoop.hive.metastore.api.StorageDescriptor()
+      tPartition.setSd(sd)
+      sd.setCols(table.schema.map(c => new FieldSchema(c.name, c.hiveType, c.comment)))
+
+      sd.setLocation(p.storage.location)
+      sd.setInputFormat(p.storage.inputFormat)
+      sd.setOutputFormat(p.storage.outputFormat)
+
+      val serdeInfo = new org.apache.hadoop.hive.metastore.api.SerDeInfo
+      sd.setSerdeInfo(serdeInfo)
+      serdeInfo.setSerializationLib(p.storage.serde)
+
+      val serdeParameters = new java.util.HashMap[String, String]()
+      serdeInfo.setParameters(serdeParameters)
+      table.serdeProperties.foreach { case (k, v) => serdeParameters.put(k, v) }
+      p.storage.serdeProperties.foreach { case (k, v) => serdeParameters.put(k, v) }
+
+      new Partition(hiveQlTable, tPartition)
+    }
+  }
+
   /** Only compare database and tablename, not alias. */
   override def sameResult(plan: LogicalPlan): Boolean = {
     plan match {

sql/hive/src/main/scala/org/apache/spark/sql/hive/HiveShim.scala

@@ -27,6 +27,7 @@ import scala.reflect.ClassTag
 
 import com.esotericsoftware.kryo.Kryo
 import com.esotericsoftware.kryo.io.{Input, Output}
+
 import org.apache.hadoop.conf.Configuration
 import org.apache.hadoop.fs.Path
 import org.apache.hadoop.hive.ql.exec.{UDF, Utilities}

sql/hive/src/main/scala/org/apache/spark/sql/hive/HiveStrategies.scala

@@ -125,7 +125,7 @@ private[hive] trait HiveStrategies {
                 InterpretedPredicate.create(castedPredicate)
               }
 
-            val partitions = relation.hiveQlPartitions.filter { part =>
+            val partitions = relation.getHiveQlPartitions(pruningPredicates).filter { part =>
               val partitionValues = part.getValues
               var i = 0
               while (i < partitionValues.size()) {
@@ -213,7 +213,7 @@ private[hive] trait HiveStrategies {
           projectList,
           otherPredicates,
           identity[Seq[Expression]],
-          HiveTableScan(_, relation, pruningPredicates.reduceLeftOption(And))(hiveContext)) :: Nil
+          HiveTableScan(_, relation, pruningPredicates)(hiveContext)) :: Nil
       case _ =>
         Nil
     }

sql/hive/src/main/scala/org/apache/spark/sql/hive/client/ClientInterface.scala

@@ -21,6 +21,7 @@ import java.io.PrintStream
 import java.util.{Map => JMap}
 
 import org.apache.spark.sql.catalyst.analysis.{NoSuchDatabaseException, NoSuchTableException}
+import org.apache.spark.sql.catalyst.expressions.Expression
 
 private[hive] case class HiveDatabase(
     name: String,
@@ -71,7 +72,12 @@ private[hive] case class HiveTable(
 
   def isPartitioned: Boolean = partitionColumns.nonEmpty
 
-  def getAllPartitions: Seq[HivePartition] = client.getAllPartitions(this)
+  def getPartitions(predicates: Seq[Expression]): Seq[HivePartition] = {
+    predicates match {
+      case Nil => client.getAllPartitions(this)
+      case _ => client.getPartitionsByFilter(this, predicates)
+    }
+  }
 
   // Hive does not support backticks when passing names to the client.
   def qualifiedName: String = s"$database.$name"
@@ -132,6 +138,9 @@ private[hive] trait ClientInterface {
   /** Returns all partitions for the given table. */
   def getAllPartitions(hTable: HiveTable): Seq[HivePartition]
 
+  /** Returns partitions filtered by predicates for the given table. */
+  def getPartitionsByFilter(hTable: HiveTable, predicates: Seq[Expression]): Seq[HivePartition]
+
   /** Loads a static partition into an existing table. */
   def loadPartition(
       loadPath: String,

sql/hive/src/main/scala/org/apache/spark/sql/hive/client/ClientWrapper.scala

@@ -17,27 +17,24 @@
 
 package org.apache.spark.sql.hive.client
 
-import java.io.{BufferedReader, InputStreamReader, File, PrintStream}
-import java.net.URI
-import java.util.{ArrayList => JArrayList, Map => JMap, List => JList, Set => JSet}
+import java.io.{File, PrintStream}
+import java.util.{Map => JMap}
 import javax.annotation.concurrent.GuardedBy
 
+import org.apache.spark.sql.catalyst.expressions.Expression
 import org.apache.spark.util.CircularBuffer
 
 import scala.collection.JavaConversions._
 import scala.language.reflectiveCalls
 
 import org.apache.hadoop.fs.Path
-import org.apache.hadoop.hive.metastore.api.Database
 import org.apache.hadoop.hive.conf.HiveConf
+import org.apache.hadoop.hive.metastore.api.{Database, FieldSchema}
 import org.apache.hadoop.hive.metastore.{TableType => HTableType}
-import org.apache.hadoop.hive.metastore.api
-import org.apache.hadoop.hive.metastore.api.FieldSchema
-import org.apache.hadoop.hive.ql.metadata
 import org.apache.hadoop.hive.ql.metadata.Hive
-import org.apache.hadoop.hive.ql.session.SessionState
 import org.apache.hadoop.hive.ql.processors._
-import org.apache.hadoop.hive.ql.Driver
+import org.apache.hadoop.hive.ql.session.SessionState
+import org.apache.hadoop.hive.ql.{Driver, metadata}
 
 import org.apache.spark.Logging
 import org.apache.spark.sql.execution.QueryExecutionException
@@ -316,6 +313,13 @@ private[hive] class ClientWrapper(
     shim.getAllPartitions(client, qlTable).map(toHivePartition)
   }
 
+  override def getPartitionsByFilter(
+      hTable: HiveTable,
+      predicates: Seq[Expression]): Seq[HivePartition] = withHiveState {
+    val qlTable = toQlTable(hTable)
+    shim.getPartitionsByFilter(client, qlTable, predicates).map(toHivePartition)
+  }
+
   override def listTables(dbName: String): Seq[String] = withHiveState {
     client.getAllTables(dbName)
   }

sql/hive/src/main/scala/org/apache/spark/sql/hive/client/HiveShim.scala

@@ -31,6 +31,11 @@ import org.apache.hadoop.hive.ql.Driver
 import org.apache.hadoop.hive.ql.metadata.{Hive, Partition, Table}
 import org.apache.hadoop.hive.ql.processors.{CommandProcessor, CommandProcessorFactory}
 import org.apache.hadoop.hive.ql.session.SessionState
+import org.apache.hadoop.hive.serde.serdeConstants
+
+import org.apache.spark.Logging
+import org.apache.spark.sql.catalyst.expressions.{Expression, AttributeReference, BinaryComparison}
+import org.apache.spark.sql.types.{StringType, IntegralType}
 
 /**
  * A shim that defines the interface between ClientWrapper and the underlying Hive library used to
@@ -61,6 +66,8 @@ private[client] sealed abstract class Shim {
 
   def getAllPartitions(hive: Hive, table: Table): Seq[Partition]
 
+  def getPartitionsByFilter(hive: Hive, table: Table, predicates: Seq[Expression]): Seq[Partition]
+
   def getCommandProcessor(token: String, conf: HiveConf): CommandProcessor
 
   def getDriverResults(driver: Driver): Seq[String]
@@ -109,7 +116,7 @@ private[client] sealed abstract class Shim {
 
 }
 
-private[client] class Shim_v0_12 extends Shim {
+private[client] class Shim_v0_12 extends Shim with Logging {
 
   private lazy val startMethod =
     findStaticMethod(
@@ -196,6 +203,17 @@ private[client] class Shim_v0_12 extends Shim {
   override def getAllPartitions(hive: Hive, table: Table): Seq[Partition] =
     getAllPartitionsMethod.invoke(hive, table).asInstanceOf[JSet[Partition]].toSeq
 
+  override def getPartitionsByFilter(
+      hive: Hive,
+      table: Table,
+      predicates: Seq[Expression]): Seq[Partition] = {
+    // getPartitionsByFilter() doesn't support binary comparison ops in Hive 0.12.
+    // See HIVE-4888.
+    logDebug("Hive 0.12 doesn't support predicate pushdown to metastore. " +
+      "Please use Hive 0.13 or higher.")
+    getAllPartitions(hive, table)
+  }
+
   override def getCommandProcessor(token: String, conf: HiveConf): CommandProcessor =
     getCommandProcessorMethod.invoke(null, token, conf).asInstanceOf[CommandProcessor]
 
@@ -267,6 +285,12 @@ private[client] class Shim_v0_13 extends Shim_v0_12 {
       classOf[Hive],
       "getAllPartitionsOf",
       classOf[Table])
+  private lazy val getPartitionsByFilterMethod =
+    findMethod(
+      classOf[Hive],
+      "getPartitionsByFilter",
+      classOf[Table],
+      classOf[String])
   private lazy val getCommandProcessorMethod =
     findStaticMethod(
       classOf[CommandProcessorFactory],
@@ -288,6 +312,48 @@ private[client] class Shim_v0_13 extends Shim_v0_12 {
   override def getAllPartitions(hive: Hive, table: Table): Seq[Partition] =
     getAllPartitionsMethod.invoke(hive, table).asInstanceOf[JSet[Partition]].toSeq
 
+  override def getPartitionsByFilter(
+      hive: Hive,
+      table: Table,
+      predicates: Seq[Expression]): Seq[Partition] = {
+    // hive varchar is treated as catalyst string, but hive varchar can't be pushed down.
+    val varcharKeys = table.getPartitionKeys
+      .filter(col => col.getType.startsWith(serdeConstants.VARCHAR_TYPE_NAME))
+      .map(col => col.getName).toSet
+
+    // Hive getPartitionsByFilter() takes a string that represents partition
+    // predicates like "str_key=\"value\" and int_key=1 ..."
+    val filter = predicates.flatMap { expr =>
+      expr match {
+        case op @ BinaryComparison(lhs, rhs) => {
+          lhs match {
+            case AttributeReference(_, _, _, _) => {
+              rhs.dataType match {
+                case _: IntegralType =>
+                  Some(lhs.prettyString + op.symbol + rhs.prettyString)
+                case _: StringType if (!varcharKeys.contains(lhs.prettyString)) =>
+                  Some(lhs.prettyString + op.symbol + "\"" + rhs.prettyString + "\"")
+                case _ => None
+              }
+            }
+            case _ => None
+          }
+        }
+        case _ => None
+      }
+    }.mkString(" and ")
+
+    val partitions =
+      if (filter.isEmpty) {
+        getAllPartitionsMethod.invoke(hive, table).asInstanceOf[JSet[Partition]]
+      } else {
+        logDebug(s"Hive metastore filter is '$filter'.")
+        getPartitionsByFilterMethod.invoke(hive, table, filter).asInstanceOf[JArrayList[Partition]]
+      }
+
+    partitions.toSeq
+  }
+
   override def getCommandProcessor(token: String, conf: HiveConf): CommandProcessor =
     getCommandProcessorMethod.invoke(null, Array(token), conf).asInstanceOf[CommandProcessor]
 

sql/hive/src/main/scala/org/apache/spark/sql/hive/execution/HiveTableScan.scala

@@ -44,7 +44,7 @@ private[hive]
 case class HiveTableScan(
     requestedAttributes: Seq[Attribute],
     relation: MetastoreRelation,
-    partitionPruningPred: Option[Expression])(
+    partitionPruningPred: Seq[Expression])(
     @transient val context: HiveContext)
   extends LeafNode {
 
@@ -56,7 +56,7 @@ case class HiveTableScan(
 
   // Bind all partition key attribute references in the partition pruning predicate for later
   // evaluation.
-  private[this] val boundPruningPred = partitionPruningPred.map { pred =>
+  private[this] val boundPruningPred = partitionPruningPred.reduceLeftOption(And).map { pred =>
     require(
       pred.dataType == BooleanType,
       s"Data type of predicate $pred must be BooleanType rather than ${pred.dataType}.")
@@ -133,7 +133,8 @@ case class HiveTableScan(
   protected override def doExecute(): RDD[InternalRow] = if (!relation.hiveQlTable.isPartitioned) {
     hadoopReader.makeRDDForTable(relation.hiveQlTable)
   } else {
-    hadoopReader.makeRDDForPartitionedTable(prunePartitions(relation.hiveQlPartitions))
+    hadoopReader.makeRDDForPartitionedTable(
+      prunePartitions(relation.getHiveQlPartitions(partitionPruningPred)))
   }
 
   override def output: Seq[Attribute] = attributes

sql/hive/src/test/scala/org/apache/spark/sql/hive/client/VersionsSuite.scala

@@ -20,7 +20,9 @@ package org.apache.spark.sql.hive.client
 import java.io.File
 
 import org.apache.spark.{Logging, SparkFunSuite}
+import org.apache.spark.sql.catalyst.expressions.{NamedExpression, Literal, AttributeReference, EqualTo}
 import org.apache.spark.sql.catalyst.util.quietly
+import org.apache.spark.sql.types.IntegerType
 import org.apache.spark.util.Utils
 
 /**
@@ -151,6 +153,12 @@ class VersionsSuite extends SparkFunSuite with Logging {
       client.getAllPartitions(client.getTable("default", "src_part"))
     }
 
+    test(s"$version: getPartitionsByFilter") {
+      client.getPartitionsByFilter(client.getTable("default", "src_part"), Seq(EqualTo(
+        AttributeReference("key", IntegerType, false)(NamedExpression.newExprId),
+        Literal(1))))
+    }
+
     test(s"$version: loadPartition") {
       client.loadPartition(
         emptyDir,

sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/PruningSuite.scala

@@ -151,7 +151,7 @@ class PruningSuite extends HiveComparisonTest with BeforeAndAfter {
         case p @ HiveTableScan(columns, relation, _) =>
           val columnNames = columns.map(_.name)
           val partValues = if (relation.table.isPartitioned) {
-            p.prunePartitions(relation.hiveQlPartitions).map(_.getValues)
+            p.prunePartitions(relation.getHiveQlPartitions()).map(_.getValues)
           } else {
             Seq.empty
           }