Unify the logic for column pruning, projection, and filtering of table scans.
This removes duplicated logic, dead code and casting when planning parquet table scans and hive table scans. Other changes: - Fix tests now that we are doing a better job of column pruning (i.e., since pruning predicates are applied before we even start scanning tuples, columns required by these predicates do not need to be included in the output of the scan unless they are also included in the final output of this logical plan fragment). - Add rule to simplify trivial filters. This was required to avoid `WHERE false` from getting pushed into table scans, since `HiveTableScan` (reasonably) refuses to apply partition pruning predicates to non-partitioned tables. Author: Michael Armbrust <michael@databricks.com> Closes #213 from marmbrus/strategyCleanup and squashes the following commits: 48ce403 [Michael Armbrust] Move one more bit of parquet stuff into the core SQLContext. 834ce08 [Michael Armbrust] Address comments. 0f2c6f5 [Michael Armbrust] Unify the logic for column pruning, projection, and filtering of table scans for both Hive and Parquet relations. Fix tests now that we are doing a better job of column pruning.
This commit is contained in:
Michael Armbrust
Patrick Wendell
parent
5140598df8
commit
b637f2d91a
|
|
@ -30,6 +30,7 @@ object Optimizer extends RuleExecutor[LogicalPlan] {
|
|||
Batch("ConstantFolding", Once,
|
||||
ConstantFolding,
|
||||
BooleanSimplification,
|
||||
SimplifyFilters,
|
||||
SimplifyCasts) ::
|
||||
Batch("Filter Pushdown", Once,
|
||||
CombineFilters,
|
||||
|
|
@ -90,6 +91,22 @@ object CombineFilters extends Rule[LogicalPlan] {
|
|||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Removes filters that can be evaluated trivially. This is done either by eliding the filter for
|
||||
* cases where it will always evaluate to `true`, or substituting a dummy empty relation when the
|
||||
* filter will always evaluate to `false`.
|
||||
*/
|
||||
object SimplifyFilters extends Rule[LogicalPlan] {
|
||||
def apply(plan: LogicalPlan): LogicalPlan = plan transform {
|
||||
case Filter(Literal(true, BooleanType), child) =>
|
||||
child
|
||||
case Filter(Literal(null, _), child) =>
|
||||
LocalRelation(child.output)
|
||||
case Filter(Literal(false, BooleanType), child) =>
|
||||
LocalRelation(child.output)
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Pushes [[catalyst.plans.logical.Filter Filter]] operators through
|
||||
* [[catalyst.plans.logical.Project Project]] operators, in-lining any
|
||||
|
|
|
|||
|
|
@ -118,9 +118,47 @@ class SQLContext(@transient val sparkContext: SparkContext)
|
|||
TopK ::
|
||||
PartialAggregation ::
|
||||
SparkEquiInnerJoin ::
|
||||
ParquetOperations ::
|
||||
BasicOperators ::
|
||||
CartesianProduct ::
|
||||
BroadcastNestedLoopJoin :: Nil
|
||||
|
||||
/**
|
||||
* Used to build table scan operators where complex projection and filtering are done using
|
||||
* separate physical operators. This function returns the given scan operator with Project and
|
||||
* Filter nodes added only when needed. For example, a Project operator is only used when the
|
||||
* final desired output requires complex expressions to be evaluated or when columns can be
|
||||
* further eliminated out after filtering has been done.
|
||||
*
|
||||
* The required attributes for both filtering and expression evaluation are passed to the
|
||||
* provided `scanBuilder` function so that it can avoid unnecessary column materialization.
|
||||
*/
|
||||
def pruneFilterProject(
|
||||
projectList: Seq[NamedExpression],
|
||||
filterPredicates: Seq[Expression],
|
||||
scanBuilder: Seq[Attribute] => SparkPlan): SparkPlan = {
|
||||
|
||||
val projectSet = projectList.flatMap(_.references).toSet
|
||||
val filterSet = filterPredicates.flatMap(_.references).toSet
|
||||
val filterCondition = filterPredicates.reduceLeftOption(And)
|
||||
|
||||
// Right now we still use a projection even if the only evaluation is applying an alias
|
||||
// to a column. Since this is a no-op, it could be avoided. However, using this
|
||||
// optimization with the current implementation would change the output schema.
|
||||
// TODO: Decouple final output schema from expression evaluation so this copy can be
|
||||
// avoided safely.
|
||||
|
||||
if (projectList.toSet == projectSet && filterSet.subsetOf(projectSet)) {
|
||||
// When it is possible to just use column pruning to get the right projection and
|
||||
// when the columns of this projection are enough to evaluate all filter conditions,
|
||||
// just do a scan followed by a filter, with no extra project.
|
||||
val scan = scanBuilder(projectList.asInstanceOf[Seq[Attribute]])
|
||||
filterCondition.map(Filter(_, scan)).getOrElse(scan)
|
||||
} else {
|
||||
val scan = scanBuilder((projectSet ++ filterSet).toSeq)
|
||||
Project(projectList, filterCondition.map(Filter(_, scan)).getOrElse(scan))
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
@transient
|
||||
|
|
|
|||
|
|
@ -18,19 +18,15 @@
|
|||
package org.apache.spark.sql
|
||||
package execution
|
||||
|
||||
import org.apache.spark.SparkContext
|
||||
|
||||
import org.apache.spark.sql.catalyst.expressions._
|
||||
import org.apache.spark.sql.catalyst.planning._
|
||||
import org.apache.spark.sql.catalyst.plans._
|
||||
import org.apache.spark.sql.catalyst.plans.logical.LogicalPlan
|
||||
import org.apache.spark.sql.catalyst.plans.physical._
|
||||
import org.apache.spark.sql.parquet.ParquetRelation
|
||||
import org.apache.spark.sql.parquet.InsertIntoParquetTable
|
||||
import org.apache.spark.sql.parquet._
|
||||
|
||||
abstract class SparkStrategies extends QueryPlanner[SparkPlan] {
|
||||
|
||||
val sparkContext: SparkContext
|
||||
self: SQLContext#SparkPlanner =>
|
||||
|
||||
object SparkEquiInnerJoin extends Strategy {
|
||||
def apply(plan: LogicalPlan): Seq[SparkPlan] = plan match {
|
||||
|
|
@ -170,6 +166,25 @@ abstract class SparkStrategies extends QueryPlanner[SparkPlan] {
|
|||
}
|
||||
}
|
||||
|
||||
object ParquetOperations extends Strategy {
|
||||
def apply(plan: LogicalPlan): Seq[SparkPlan] = plan match {
|
||||
// TODO: need to support writing to other types of files. Unify the below code paths.
|
||||
case logical.WriteToFile(path, child) =>
|
||||
val relation =
|
||||
ParquetRelation.create(path, child, sparkContext.hadoopConfiguration, None)
|
||||
InsertIntoParquetTable(relation, planLater(child))(sparkContext) :: Nil
|
||||
case logical.InsertIntoTable(table: ParquetRelation, partition, child, overwrite) =>
|
||||
InsertIntoParquetTable(table, planLater(child))(sparkContext) :: Nil
|
||||
case PhysicalOperation(projectList, filters, relation: parquet.ParquetRelation) =>
|
||||
// TODO: Should be pushing down filters as well.
|
||||
pruneFilterProject(
|
||||
projectList,
|
||||
filters,
|
||||
ParquetTableScan(_, relation, None)(sparkContext)) :: Nil
|
||||
case _ => Nil
|
||||
}
|
||||
}
|
||||
|
||||
// Can we automate these 'pass through' operations?
|
||||
object BasicOperators extends Strategy {
|
||||
// TODO: Set
|
||||
|
|
@ -185,14 +200,6 @@ abstract class SparkStrategies extends QueryPlanner[SparkPlan] {
|
|||
// This sort only sorts tuples within a partition. Its requiredDistribution will be
|
||||
// an UnspecifiedDistribution.
|
||||
execution.Sort(sortExprs, global = false, planLater(child)) :: Nil
|
||||
case logical.Project(projectList, r: ParquetRelation)
|
||||
if projectList.forall(_.isInstanceOf[Attribute]) =>
|
||||
|
||||
// simple projection of data loaded from Parquet file
|
||||
parquet.ParquetTableScan(
|
||||
projectList.asInstanceOf[Seq[Attribute]],
|
||||
r,
|
||||
None)(sparkContext) :: Nil
|
||||
case logical.Project(projectList, child) =>
|
||||
execution.Project(projectList, planLater(child)) :: Nil
|
||||
case logical.Filter(condition, child) =>
|
||||
|
|
@ -216,12 +223,6 @@ abstract class SparkStrategies extends QueryPlanner[SparkPlan] {
|
|||
execution.ExistingRdd(Nil, singleRowRdd) :: Nil
|
||||
case logical.Repartition(expressions, child) =>
|
||||
execution.Exchange(HashPartitioning(expressions, numPartitions), planLater(child)) :: Nil
|
||||
case logical.WriteToFile(path, child) =>
|
||||
val relation =
|
||||
ParquetRelation.create(path, child, sparkContext.hadoopConfiguration, None)
|
||||
InsertIntoParquetTable(relation, planLater(child))(sparkContext) :: Nil
|
||||
case p: parquet.ParquetRelation =>
|
||||
parquet.ParquetTableScan(p.output, p, None)(sparkContext) :: Nil
|
||||
case SparkLogicalPlan(existingPlan) => existingPlan :: Nil
|
||||
case _ => Nil
|
||||
}
|
||||
|
|
|
|||
|
|
@ -133,8 +133,6 @@ class HiveContext(sc: SparkContext) extends SQLContext(sc) {
|
|||
results
|
||||
}
|
||||
|
||||
// TODO: Move this.
|
||||
|
||||
SessionState.start(sessionState)
|
||||
|
||||
/**
|
||||
|
|
@ -191,8 +189,7 @@ class HiveContext(sc: SparkContext) extends SQLContext(sc) {
|
|||
|
||||
override val strategies: Seq[Strategy] = Seq(
|
||||
TopK,
|
||||
ColumnPrunings,
|
||||
PartitionPrunings,
|
||||
ParquetOperations,
|
||||
HiveTableScans,
|
||||
DataSinks,
|
||||
Scripts,
|
||||
|
|
@ -217,7 +214,6 @@ class HiveContext(sc: SparkContext) extends SQLContext(sc) {
|
|||
override lazy val optimizedPlan =
|
||||
optimizer(catalog.PreInsertionCasts(catalog.CreateTables(analyzed)))
|
||||
|
||||
// TODO: We are loosing schema here.
|
||||
override lazy val toRdd: RDD[Row] =
|
||||
analyzed match {
|
||||
case NativeCommand(cmd) =>
|
||||
|
|
|
|||
|
|
@ -21,9 +21,8 @@ package hive
|
|||
import org.apache.spark.sql.catalyst.expressions._
|
||||
import org.apache.spark.sql.catalyst.planning._
|
||||
import org.apache.spark.sql.catalyst.plans._
|
||||
import org.apache.spark.sql.catalyst.plans.logical.{BaseRelation, LogicalPlan}
|
||||
import org.apache.spark.sql.catalyst.plans.logical.LogicalPlan
|
||||
import org.apache.spark.sql.execution._
|
||||
import org.apache.spark.sql.parquet.{InsertIntoParquetTable, ParquetRelation, ParquetTableScan}
|
||||
|
||||
trait HiveStrategies {
|
||||
// Possibly being too clever with types here... or not clever enough.
|
||||
|
|
@ -43,121 +42,31 @@ trait HiveStrategies {
|
|||
def apply(plan: LogicalPlan): Seq[SparkPlan] = plan match {
|
||||
case logical.InsertIntoTable(table: MetastoreRelation, partition, child, overwrite) =>
|
||||
InsertIntoHiveTable(table, partition, planLater(child), overwrite)(hiveContext) :: Nil
|
||||
case logical.InsertIntoTable(table: ParquetRelation, partition, child, overwrite) =>
|
||||
InsertIntoParquetTable(table, planLater(child))(hiveContext.sparkContext) :: Nil
|
||||
case _ => Nil
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Retrieves data using a HiveTableScan. Partition pruning predicates are also detected and
|
||||
* applied.
|
||||
*/
|
||||
object HiveTableScans extends Strategy {
|
||||
def apply(plan: LogicalPlan): Seq[SparkPlan] = plan match {
|
||||
// Push attributes into table scan when possible.
|
||||
case p @ logical.Project(projectList, m: MetastoreRelation) if isSimpleProject(projectList) =>
|
||||
HiveTableScan(projectList.asInstanceOf[Seq[Attribute]], m, None)(hiveContext) :: Nil
|
||||
case m: MetastoreRelation =>
|
||||
HiveTableScan(m.output, m, None)(hiveContext) :: Nil
|
||||
case _ => Nil
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* A strategy used to detect filtering predicates on top of a partitioned relation to help
|
||||
* partition pruning.
|
||||
*
|
||||
* This strategy itself doesn't perform partition pruning, it just collects and combines all the
|
||||
* partition pruning predicates and pass them down to the underlying [[HiveTableScan]] operator,
|
||||
* which does the actual pruning work.
|
||||
*/
|
||||
object PartitionPrunings extends Strategy {
|
||||
def apply(plan: LogicalPlan): Seq[SparkPlan] = plan match {
|
||||
case p @ FilteredOperation(predicates, relation: MetastoreRelation)
|
||||
if relation.isPartitioned =>
|
||||
|
||||
case PhysicalOperation(projectList, predicates, relation: MetastoreRelation) =>
|
||||
// Filter out all predicates that only deal with partition keys, these are given to the
|
||||
// hive table scan operator to be used for partition pruning.
|
||||
val partitionKeyIds = relation.partitionKeys.map(_.exprId).toSet
|
||||
|
||||
// Filter out all predicates that only deal with partition keys
|
||||
val (pruningPredicates, otherPredicates) = predicates.partition {
|
||||
_.references.map(_.exprId).subsetOf(partitionKeyIds)
|
||||
|
||||
}
|
||||
|
||||
val scan = HiveTableScan(
|
||||
relation.output, relation, pruningPredicates.reduceLeftOption(And))(hiveContext)
|
||||
|
||||
otherPredicates
|
||||
.reduceLeftOption(And)
|
||||
.map(Filter(_, scan))
|
||||
.getOrElse(scan) :: Nil
|
||||
|
||||
pruneFilterProject(
|
||||
projectList,
|
||||
otherPredicates,
|
||||
HiveTableScan(_, relation, pruningPredicates.reduceLeftOption(And))(hiveContext)) :: Nil
|
||||
case _ =>
|
||||
Nil
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* A strategy that detects projects and filters over some relation and applies column pruning if
|
||||
* possible. Partition pruning is applied first if the relation is partitioned.
|
||||
*/
|
||||
object ColumnPrunings extends Strategy {
|
||||
def apply(plan: LogicalPlan): Seq[SparkPlan] = plan match {
|
||||
// TODO(andre): the current mix of HiveRelation and ParquetRelation
|
||||
// here appears artificial; try to refactor to break it into two
|
||||
case PhysicalOperation(projectList, predicates, relation: BaseRelation) =>
|
||||
val predicateOpt = predicates.reduceOption(And)
|
||||
val predicateRefs = predicateOpt.map(_.references).getOrElse(Set.empty)
|
||||
val projectRefs = projectList.flatMap(_.references)
|
||||
|
||||
// To figure out what columns to preserve after column pruning, we need to consider:
|
||||
//
|
||||
// 1. Columns referenced by the project list (order preserved)
|
||||
// 2. Columns referenced by filtering predicates but not by project list
|
||||
// 3. Relation output
|
||||
//
|
||||
// Then the final result is ((1 union 2) intersect 3)
|
||||
val prunedCols = (projectRefs ++ (predicateRefs -- projectRefs)).intersect(relation.output)
|
||||
|
||||
val filteredScans =
|
||||
if (relation.isPartitioned) { // from here on relation must be a [[MetaStoreRelation]]
|
||||
// Applies partition pruning first for partitioned table
|
||||
val filteredRelation = predicateOpt.map(logical.Filter(_, relation)).getOrElse(relation)
|
||||
PartitionPrunings(filteredRelation).view.map(_.transform {
|
||||
case scan: HiveTableScan =>
|
||||
scan.copy(attributes = prunedCols)(hiveContext)
|
||||
})
|
||||
} else {
|
||||
val scan = relation match {
|
||||
case MetastoreRelation(_, _, _) => {
|
||||
HiveTableScan(
|
||||
prunedCols,
|
||||
relation.asInstanceOf[MetastoreRelation],
|
||||
None)(hiveContext)
|
||||
}
|
||||
case ParquetRelation(_, _) => {
|
||||
ParquetTableScan(
|
||||
relation.output,
|
||||
relation.asInstanceOf[ParquetRelation],
|
||||
None)(hiveContext.sparkContext)
|
||||
.pruneColumns(prunedCols)
|
||||
}
|
||||
}
|
||||
predicateOpt.map(execution.Filter(_, scan)).getOrElse(scan) :: Nil
|
||||
}
|
||||
|
||||
if (isSimpleProject(projectList) && prunedCols == projectRefs) {
|
||||
filteredScans
|
||||
} else {
|
||||
filteredScans.view.map(execution.Project(projectList, _))
|
||||
}
|
||||
|
||||
case _ =>
|
||||
Nil
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Returns true if `projectList` only performs column pruning and does not evaluate other
|
||||
* complex expressions.
|
||||
*/
|
||||
def isSimpleProject(projectList: Seq[NamedExpression]) = {
|
||||
projectList.forall(_.isInstanceOf[Attribute])
|
||||
}
|
||||
}
|
||||
|
|
|
|||
|
|
@ -33,7 +33,7 @@ class PruningSuite extends HiveComparisonTest {
|
|||
createPruningTest("Column pruning: with partitioned table",
|
||||
"SELECT key FROM srcpart WHERE ds = '2008-04-08' LIMIT 3",
|
||||
Seq("key"),
|
||||
Seq("key", "ds"),
|
||||
Seq("key"),
|
||||
Seq(
|
||||
Seq("2008-04-08", "11"),
|
||||
Seq("2008-04-08", "12")))
|
||||
|
|
@ -97,7 +97,7 @@ class PruningSuite extends HiveComparisonTest {
|
|||
createPruningTest("Partition pruning: with filter on string partition key",
|
||||
"SELECT value, hr FROM srcpart1 WHERE ds = '2008-04-08'",
|
||||
Seq("value", "hr"),
|
||||
Seq("value", "hr", "ds"),
|
||||
Seq("value", "hr"),
|
||||
Seq(
|
||||
Seq("2008-04-08", "11"),
|
||||
Seq("2008-04-08", "12")))
|
||||
|
|
@ -113,14 +113,14 @@ class PruningSuite extends HiveComparisonTest {
|
|||
createPruningTest("Partition pruning: left only 1 partition",
|
||||
"SELECT value, hr FROM srcpart1 WHERE ds = '2008-04-08' AND hr < 12",
|
||||
Seq("value", "hr"),
|
||||
Seq("value", "hr", "ds"),
|
||||
Seq("value", "hr"),
|
||||
Seq(
|
||||
Seq("2008-04-08", "11")))
|
||||
|
||||
createPruningTest("Partition pruning: all partitions pruned",
|
||||
"SELECT value, hr FROM srcpart1 WHERE ds = '2014-01-27' AND hr = 11",
|
||||
Seq("value", "hr"),
|
||||
Seq("value", "hr", "ds"),
|
||||
Seq("value", "hr"),
|
||||
Seq.empty)
|
||||
|
||||
createPruningTest("Partition pruning: pruning with both column key and partition key",
|
||||
|
|
@ -147,8 +147,8 @@ class PruningSuite extends HiveComparisonTest {
|
|||
(columnNames, partValues)
|
||||
}.head
|
||||
|
||||
assert(actualOutputColumns sameElements expectedOutputColumns, "Output columns mismatch")
|
||||
assert(actualScannedColumns sameElements expectedScannedColumns, "Scanned columns mismatch")
|
||||
assert(actualOutputColumns === expectedOutputColumns, "Output columns mismatch")
|
||||
assert(actualScannedColumns === expectedScannedColumns, "Scanned columns mismatch")
|
||||
|
||||
assert(
|
||||
actualPartValues.length === expectedPartValues.length,
|
||||
|
|
|
|||
Loading…
Reference in a new issue