Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[SPARK-33427][SQL] Add subexpression elimination for interpreted expression evaluation #30341

Closed
wants to merge 10 commits into from
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,7 @@ import org.apache.spark.SparkException
import org.apache.spark.sql.catalyst.InternalRow
import org.apache.spark.sql.catalyst.expressions.codegen.{UnsafeArrayWriter, UnsafeRowWriter, UnsafeWriter}
import org.apache.spark.sql.catalyst.util.ArrayData
import org.apache.spark.sql.internal.SQLConf
import org.apache.spark.sql.types.{UserDefinedType, _}
import org.apache.spark.unsafe.Platform

Expand All @@ -33,6 +34,15 @@ import org.apache.spark.unsafe.Platform
class InterpretedUnsafeProjection(expressions: Array[Expression]) extends UnsafeProjection {
import InterpretedUnsafeProjection._

private[this] val subExprEliminationEnabled = SQLConf.get.subexpressionEliminationEnabled
private[this] lazy val runtime =
new SubExprEvaluationRuntime(SQLConf.get.subexpressionEliminationCacheMaxEntries)
private[this] val exprs = if (subExprEliminationEnabled) {
runtime.proxyExpressions(expressions)
} else {
expressions.toSeq
}

/** Number of (top level) fields in the resulting row. */
private[this] val numFields = expressions.length

Expand Down Expand Up @@ -63,17 +73,21 @@ class InterpretedUnsafeProjection(expressions: Array[Expression]) extends Unsafe
}

override def initialize(partitionIndex: Int): Unit = {
expressions.foreach(_.foreach {
exprs.foreach(_.foreach {
case n: Nondeterministic => n.initialize(partitionIndex)
case _ =>
})
}

override def apply(row: InternalRow): UnsafeRow = {
if (subExprEliminationEnabled) {
runtime.setInput(row)
}

// Put the expression results in the intermediate row.
var i = 0
while (i < numFields) {
values(i) = expressions(i).eval(row)
values(i) = exprs(i).eval(row)
i += 1
}

Expand Down
Original file line number Diff line number Diff line change
@@ -0,0 +1,126 @@
/*
* Licensed to the Apache Software Foundation (ASF) under one or more
* contributor license agreements. See the NOTICE file distributed with
* this work for additional information regarding copyright ownership.
* The ASF licenses this file to You under the Apache License, Version 2.0
* (the "License"); you may not use this file except in compliance with
* the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package org.apache.spark.sql.catalyst.expressions

import scala.collection.mutable

import com.google.common.cache.{CacheBuilder, CacheLoader, LoadingCache}
import com.google.common.util.concurrent.{ExecutionError, UncheckedExecutionException}

import org.apache.spark.sql.catalyst.InternalRow
import org.apache.spark.sql.catalyst.expressions.codegen.{CodegenContext, ExprCode}
import org.apache.spark.sql.types.DataType

/**
* This class helps subexpression elimination for interpreted evaluation
* such as `InterpretedUnsafeProjection`. It maintains an evaluation cache.
* This class wraps `ExpressionProxy` around given expressions. The `ExpressionProxy`
* intercepts expression evaluation and loads from the cache first.
*/
class SubExprEvaluationRuntime(cacheMaxEntries: Int) {

private[sql] val cache: LoadingCache[ExpressionProxy, ResultProxy] = CacheBuilder.newBuilder()
.maximumSize(cacheMaxEntries)
.build(
new CacheLoader[ExpressionProxy, ResultProxy]() {
override def load(expr: ExpressionProxy): ResultProxy = {
ResultProxy(expr.proxyEval(currentInput))
}
})

private var currentInput: InternalRow = null

def getEval(proxy: ExpressionProxy): Any = try {
cache.get(proxy).result
} catch {
// Cache.get() may wrap the original exception. See the following URL
// http://google.github.io/guava/releases/14.0/api/docs/com/google/common/cache/
// Cache.html#get(K,%20java.util.concurrent.Callable)
case e @ (_: UncheckedExecutionException | _: ExecutionError) =>
throw e.getCause
}

/**
* Sets given input row as current row for evaluating expressions. This cleans up the cache
* too as new input comes.
*/
def setInput(input: InternalRow = null): Unit = {
currentInput = input
cache.invalidateAll()
}

/**
* Recursively replaces expression with its proxy expression in `proxyMap`.
*/
private def replaceWithProxy(
expr: Expression,
proxyMap: Map[Expression, ExpressionProxy]): Expression = {
proxyMap.getOrElse(expr, expr.mapChildren(replaceWithProxy(_, proxyMap)))
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is a top-down traverse, which means we will recursively replace expr with proxy even for ExpressionProxy. Is it expected?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Once we replace one expression with ExpressionProxy, we stop traversing down. We only traverse down to children if cannot find current expression in proxyMap. Is this for your question?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ah sorry I misread the code. You are right.

}

/**
* Finds subexpressions and wraps them with `ExpressionProxy`.
*/
def proxyExpressions(expressions: Seq[Expression]): Seq[Expression] = {
val equivalentExpressions: EquivalentExpressions = new EquivalentExpressions

expressions.foreach(equivalentExpressions.addExprTree(_))

val proxyMap = mutable.Map.empty[Expression, ExpressionProxy]
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

is it OK to use a simple map here? Two expressions may not equal to each other even if they semantically equal.

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

For semantically equal exprs, we put a pair of expr -> proxy into the map and note the proxy is the same. So later we traverse down into expressions, we look at the map. We don't do semantically comparing when looking at this map.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ah I get it now. Seems we can use IdentityHashMap to be more explicit.


val commonExprs = equivalentExpressions.getAllEquivalentExprs.filter(_.size > 1)
commonExprs.foreach { e =>
val expr = e.head
val proxy = ExpressionProxy(expr, this)

proxyMap ++= e.map(_ -> proxy).toMap
}

// Only adding proxy if we find subexpressions.
if (proxyMap.nonEmpty) {
expressions.map(replaceWithProxy(_, proxyMap.toMap))
} else {
expressions
}
}
}

/**
* A proxy for an catalyst `Expression`. Given a runtime object `SubExprEvaluationRuntime`,
* when this is asked to evaluate, it will load from the evaluation cache in the runtime first.
*/
case class ExpressionProxy(
cloud-fan marked this conversation as resolved.
Show resolved Hide resolved
child: Expression,
runtime: SubExprEvaluationRuntime) extends Expression {

final override def dataType: DataType = child.dataType
final override def nullable: Boolean = child.nullable
final override def children: Seq[Expression] = child :: Nil

// `ExpressionProxy` is for interpreted expression evaluation only. So cannot `doGenCode`.
final override protected def doGenCode(ctx: CodegenContext, ev: ExprCode): ExprCode =
throw new UnsupportedOperationException(s"Cannot generate code for expression: $this")

def proxyEval(input: InternalRow = null): Any = child.eval(input)

override def eval(input: InternalRow = null): Any = runtime.getEval(this)
}

/**
* A simple wrapper for holding `Any` in the cache of `SubExprEvaluationRuntime`.
*/
case class ResultProxy(result: Any)
Original file line number Diff line number Diff line change
Expand Up @@ -539,6 +539,15 @@ object SQLConf {
.booleanConf
.createWithDefault(true)

val SUBEXPRESSION_ELIMINATION_CACHE_MAX_ENTRIES =
buildConf("spark.sql.subexpressionElimination.cache.maxEntries")
.internal()
.doc("The maximum entries of the cache used for interpreted subexpression elimination.")
.version("3.1.0")
.intConf
.checkValue(_ >= 0, "The maximum must not be negative")
.createWithDefault(100)
dongjoon-hyun marked this conversation as resolved.
Show resolved Hide resolved

val CASE_SENSITIVE = buildConf("spark.sql.caseSensitive")
.internal()
.doc("Whether the query analyzer should be case sensitive or not. " +
Expand Down Expand Up @@ -3214,6 +3223,9 @@ class SQLConf extends Serializable with Logging {
def subexpressionEliminationEnabled: Boolean =
getConf(SUBEXPRESSION_ELIMINATION_ENABLED)

def subexpressionEliminationCacheMaxEntries: Int =
getConf(SUBEXPRESSION_ELIMINATION_CACHE_MAX_ENTRIES)

def autoBroadcastJoinThreshold: Long = getConf(AUTO_BROADCASTJOIN_THRESHOLD)

def limitScaleUpFactor: Int = getConf(LIMIT_SCALE_UP_FACTOR)
Expand Down
Original file line number Diff line number Diff line change
@@ -0,0 +1,99 @@
/*
* Licensed to the Apache Software Foundation (ASF) under one or more
* contributor license agreements. See the NOTICE file distributed with
* this work for additional information regarding copyright ownership.
* The ASF licenses this file to You under the Apache License, Version 2.0
* (the "License"); you may not use this file except in compliance with
* the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package org.apache.spark.sql.catalyst.expressions

import org.apache.spark.SparkFunSuite

class SubExprEvaluationRuntimeSuite extends SparkFunSuite {

test("Evaluate ExpressionProxy should create cached result") {
val runtime = new SubExprEvaluationRuntime(1)
val proxy = ExpressionProxy(Literal(1), runtime)
assert(runtime.cache.size() == 0)
proxy.eval()
assert(runtime.cache.size() == 1)
assert(runtime.cache.get(proxy) == ResultProxy(1))
}

test("SubExprEvaluationRuntime cannot exceed configured max entries") {
val runtime = new SubExprEvaluationRuntime(2)
assert(runtime.cache.size() == 0)

val proxy1 = ExpressionProxy(Literal(1), runtime)
proxy1.eval()
assert(runtime.cache.size() == 1)
assert(runtime.cache.get(proxy1) == ResultProxy(1))

val proxy2 = ExpressionProxy(Literal(2), runtime)
proxy2.eval()
assert(runtime.cache.size() == 2)
assert(runtime.cache.get(proxy2) == ResultProxy(2))

val proxy3 = ExpressionProxy(Literal(3), runtime)
proxy3.eval()
assert(runtime.cache.size() == 2)
assert(runtime.cache.get(proxy3) == ResultProxy(3))
}

test("setInput should empty cached result") {
val runtime = new SubExprEvaluationRuntime(2)
val proxy1 = ExpressionProxy(Literal(1), runtime)
assert(runtime.cache.size() == 0)
proxy1.eval()
assert(runtime.cache.size() == 1)
assert(runtime.cache.get(proxy1) == ResultProxy(1))

val proxy2 = ExpressionProxy(Literal(2), runtime)
proxy2.eval()
assert(runtime.cache.size() == 2)
assert(runtime.cache.get(proxy2) == ResultProxy(2))

runtime.setInput()
assert(runtime.cache.size() == 0)
}

test("Wrap ExpressionProxy on subexpressions") {
val runtime = new SubExprEvaluationRuntime(1)

val one = Literal(1)
val two = Literal(2)
val mul = Multiply(one, two)
val mul2 = Multiply(mul, mul)
val sqrt = Sqrt(mul2)
val sum = Add(mul2, sqrt)

// ( (one * two) * (one * two) ) + sqrt( (one * two) * (one * two) )
val proxyExpressions = runtime.proxyExpressions(Seq(sum))
val proxys = proxyExpressions.flatMap(_.collect {
case p: ExpressionProxy => p
})
// ( (one * two) * (one * two) )
assert(proxys.size == 2)
val expected = ExpressionProxy(mul2, runtime)
assert(proxys.head == expected)
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

should this be proxys.forall(_ == expected)?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

yeah, you're right.

}

test("ExpressionProxy won't be on non deterministic") {
val runtime = new SubExprEvaluationRuntime(1)

val sum = Add(Rand(0), Rand(0))
val proxys = runtime.proxyExpressions(Seq(sum, sum)).flatMap(_.collect {
case p: ExpressionProxy => p
})
assert(proxys.isEmpty)
}
}
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

can we test attributes?

val attr1 = AttributeReference("a", ...)
val attr2 = attr1.withName("A")

To make sure 2 semantically-equal attributes can be optimized.

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

EquivalentExpressions skips for LeafExpression. So attributes won't be counted for subexpression.