Skip to content

[SPARK-39312][SQL] Use parquet native In predicate for in filter push down #36696

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.

Sign up for GitHub

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

Closed
huaxingao wants to merge 9 commits into from
Closed

[SPARK-39312][SQL] Use parquet native In predicate for in filter push down #36696

Show file tree
Hide file tree
Changes from 6 commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
b581aab
Use parquet native In predicate for filer IN
huaxingao May 27, 2022
4d00b20
only use parquet In predicate when the number of values in the In set…
huaxingao May 29, 2022
6773193
remove extra space
huaxingao May 29, 2022
25ceda8
address comments
huaxingao Jun 1, 2022
770a4fd
only push down in predicate when canPartialPushDownConjuncts is true
huaxingao Jun 14, 2022
08efa3c
not push down null
huaxingao Jul 11, 2022
485112e
fix test failure
huaxingao Jul 11, 2022
910328a
update benchmark results
huaxingao Jul 12, 2022
17d8ee7
Trigger Build
huaxingao Jul 12, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
151 changes: 80 additions & 71 deletions ...re/src/main/scala/org/apache/spark/sql/execution/datasources/parquet/ParquetFilters.scala
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,11 +21,11 @@ import java.lang.{Boolean => JBoolean, Double => JDouble, Float => JFloat, Long
import java.math.{BigDecimal => JBigDecimal}
import java.sql.{Date, Timestamp}
import java.time.{Duration, Instant, LocalDate, Period}
import java.util.HashSet
import java.util.Locale

import scala.collection.JavaConverters.asScalaBufferConverter

import org.apache.parquet.column.statistics.{Statistics => ParquetStatistics}
import org.apache.parquet.filter2.predicate._
import org.apache.parquet.filter2.predicate.SparkFilterApi._
import org.apache.parquet.io.api.Binary
Expand Down Expand Up @@ -444,94 +444,106 @@ class ParquetFilters(
}

private val makeInPredicate:
PartialFunction[ParquetSchemaType,
(Array[String], Array[Any], ParquetStatistics[_]) => FilterPredicate] = {
PartialFunction[ParquetSchemaType, (Array[String], Array[Any]) => FilterPredicate] = {

case ParquetByteType | ParquetShortType | ParquetIntegerType =>
(n: Array[String], v: Array[Any], statistics: ParquetStatistics[_]) =>
v.map(toIntValue(_).toInt).foreach(statistics.updateStats)
FilterApi.and(
FilterApi.gtEq(intColumn(n), statistics.genericGetMin().asInstanceOf[Integer]),
FilterApi.ltEq(intColumn(n), statistics.genericGetMax().asInstanceOf[Integer]))
(n: Array[String], values: Array[Any]) =>
val set = new HashSet[Integer]()
for (value <- values) {
set.add(toIntValue(value))
}
FilterApi.in(intColumn(n), set)

case ParquetLongType =>
(n: Array[String], v: Array[Any], statistics: ParquetStatistics[_]) =>
v.map(toLongValue).foreach(statistics.updateStats(_))
FilterApi.and(
FilterApi.gtEq(longColumn(n), statistics.genericGetMin().asInstanceOf[JLong]),
FilterApi.ltEq(longColumn(n), statistics.genericGetMax().asInstanceOf[JLong]))
(n: Array[String], values: Array[Any]) =>
val set = new HashSet[JLong]()
for (value <- values) {
set.add(toLongValue(value))
}
FilterApi.in(longColumn(n), set)

case ParquetFloatType =>
(n: Array[String], v: Array[Any], statistics: ParquetStatistics[_]) =>
v.map(_.asInstanceOf[JFloat]).foreach(statistics.updateStats(_))
FilterApi.and(
FilterApi.gtEq(floatColumn(n), statistics.genericGetMin().asInstanceOf[JFloat]),
FilterApi.ltEq(floatColumn(n), statistics.genericGetMax().asInstanceOf[JFloat]))
(n: Array[String], values: Array[Any]) =>
val set = new HashSet[JFloat]()
for (value <- values) {
set.add(value.asInstanceOf[JFloat])
}
FilterApi.in(floatColumn(n), set)

case ParquetDoubleType =>
(n: Array[String], v: Array[Any], statistics: ParquetStatistics[_]) =>
v.map(_.asInstanceOf[JDouble]).foreach(statistics.updateStats(_))
FilterApi.and(
FilterApi.gtEq(doubleColumn(n), statistics.genericGetMin().asInstanceOf[JDouble]),
FilterApi.ltEq(doubleColumn(n), statistics.genericGetMax().asInstanceOf[JDouble]))
(n: Array[String], values: Array[Any]) =>
val set = new HashSet[JDouble]()
for (value <- values) {
set.add(value.asInstanceOf[JDouble])
}
FilterApi.in(doubleColumn(n), set)

case ParquetStringType =>
(n: Array[String], v: Array[Any], statistics: ParquetStatistics[_]) =>
v.map(s => Binary.fromString(s.asInstanceOf[String])).foreach(statistics.updateStats)
FilterApi.and(
FilterApi.gtEq(binaryColumn(n), statistics.genericGetMin().asInstanceOf[Binary]),
FilterApi.ltEq(binaryColumn(n), statistics.genericGetMax().asInstanceOf[Binary]))
(n: Array[String], values: Array[Any]) =>
val set = new HashSet[Binary]()
for (value <- values) {
set.add(Option(value).map(s => Binary.fromString(s.asInstanceOf[String])).orNull)
}
FilterApi.in(binaryColumn(n), set)

case ParquetBinaryType =>
(n: Array[String], v: Array[Any], statistics: ParquetStatistics[_]) =>
v.map(b => Binary.fromReusedByteArray(b.asInstanceOf[Array[Byte]]))
.foreach(statistics.updateStats)
FilterApi.and(
FilterApi.gtEq(binaryColumn(n), statistics.genericGetMin().asInstanceOf[Binary]),
FilterApi.ltEq(binaryColumn(n), statistics.genericGetMax().asInstanceOf[Binary]))
(n: Array[String], values: Array[Any]) =>
val set = new HashSet[Binary]()
for (value <- values) {
set.add(Option(value)
.map(b => Binary.fromReusedByteArray(b.asInstanceOf[Array[Byte]])).orNull)
}
FilterApi.in(binaryColumn(n), set)

case ParquetDateType if pushDownDate =>
(n: Array[String], v: Array[Any], statistics: ParquetStatistics[_]) =>
v.map(dateToDays).map(_.asInstanceOf[Integer]).foreach(statistics.updateStats(_))
FilterApi.and(
FilterApi.gtEq(intColumn(n), statistics.genericGetMin().asInstanceOf[Integer]),
FilterApi.ltEq(intColumn(n), statistics.genericGetMax().asInstanceOf[Integer]))
(n: Array[String], values: Array[Any]) =>
val set = new HashSet[Integer]()
for (value <- values) {
set.add(Option(value).map(date => dateToDays(date).asInstanceOf[Integer]).orNull)
}
FilterApi.in(intColumn(n), set)

case ParquetTimestampMicrosType if pushDownTimestamp =>
(n: Array[String], v: Array[Any], statistics: ParquetStatistics[_]) =>
v.map(timestampToMicros).foreach(statistics.updateStats(_))
FilterApi.and(
FilterApi.gtEq(longColumn(n), statistics.genericGetMin().asInstanceOf[JLong]),
FilterApi.ltEq(longColumn(n), statistics.genericGetMax().asInstanceOf[JLong]))
(n: Array[String], values: Array[Any]) =>
val set = new HashSet[JLong]()
for (value <- values) {
set.add(Option(value).map(timestampToMicros).orNull)
}
FilterApi.in(longColumn(n), set)

case ParquetTimestampMillisType if pushDownTimestamp =>
(n: Array[String], v: Array[Any], statistics: ParquetStatistics[_]) =>
v.map(timestampToMillis).foreach(statistics.updateStats(_))
FilterApi.and(
FilterApi.gtEq(longColumn(n), statistics.genericGetMin().asInstanceOf[JLong]),
FilterApi.ltEq(longColumn(n), statistics.genericGetMax().asInstanceOf[JLong]))
(n: Array[String], values: Array[Any]) =>
val set = new HashSet[JLong]()
for (value <- values) {
set.add(Option(value).map(timestampToMillis).orNull)
}
FilterApi.in(longColumn(n), set)

case ParquetSchemaType(_: DecimalLogicalTypeAnnotation, INT32, _) if pushDownDecimal =>
(n: Array[String], v: Array[Any], statistics: ParquetStatistics[_]) =>
v.map(_.asInstanceOf[JBigDecimal]).map(decimalToInt32).foreach(statistics.updateStats(_))
FilterApi.and(
FilterApi.gtEq(intColumn(n), statistics.genericGetMin().asInstanceOf[Integer]),
FilterApi.ltEq(intColumn(n), statistics.genericGetMax().asInstanceOf[Integer]))
(n: Array[String], values: Array[Any]) =>
val set = new HashSet[Integer]()
for (value <- values) {
set.add(Option(value).map(d => decimalToInt32(d.asInstanceOf[JBigDecimal])).orNull)
}
FilterApi.in(intColumn(n), set)

case ParquetSchemaType(_: DecimalLogicalTypeAnnotation, INT64, _) if pushDownDecimal =>
(n: Array[String], v: Array[Any], statistics: ParquetStatistics[_]) =>
v.map(_.asInstanceOf[JBigDecimal]).map(decimalToInt64).foreach(statistics.updateStats(_))
FilterApi.and(
FilterApi.gtEq(longColumn(n), statistics.genericGetMin().asInstanceOf[JLong]),
FilterApi.ltEq(longColumn(n), statistics.genericGetMax().asInstanceOf[JLong]))
(n: Array[String], values: Array[Any]) =>
val set = new HashSet[JLong]()
for (value <- values) {
set.add(Option(value).map(d => decimalToInt64(d.asInstanceOf[JBigDecimal])).orNull)
}
FilterApi.in(longColumn(n), set)

case ParquetSchemaType(_: DecimalLogicalTypeAnnotation, FIXED_LEN_BYTE_ARRAY, length)
if pushDownDecimal =>
(path: Array[String], v: Array[Any], statistics: ParquetStatistics[_]) =>
v.map(d => decimalToByteArray(d.asInstanceOf[JBigDecimal], length))
.foreach(statistics.updateStats)
FilterApi.and(
FilterApi.gtEq(binaryColumn(path), statistics.genericGetMin().asInstanceOf[Binary]),
FilterApi.ltEq(binaryColumn(path), statistics.genericGetMax().asInstanceOf[Binary]))
if pushDownDecimal =>
(n: Array[String], values: Array[Any]) =>
val set = new HashSet[Binary]()
for (value <- values) {
set.add(Option(value)
.map(d => decimalToByteArray(d.asInstanceOf[JBigDecimal], length)).orNull)
}
FilterApi.in(binaryColumn(n), set)
}

// Returns filters that can be pushed down when reading Parquet files.
Expand Down Expand Up @@ -736,15 +748,12 @@ class ParquetFilters(
makeEq.lift(fieldType).map(_(fieldNames, v))
}.reduceLeftOption(FilterApi.or)
} else if (canPartialPushDownConjuncts) {
val primitiveType = schema.getColumnDescription(fieldNames).getPrimitiveType
val statistics: ParquetStatistics[_] = ParquetStatistics.createStats(primitiveType)
if (values.contains(null)) {
Seq(makeEq.lift(fieldType).map(_(fieldNames, null)),
makeInPredicate.lift(fieldType)
.map(_(fieldNames, values.filter(_ != null), statistics))
makeInPredicate.lift(fieldType).map(_(fieldNames, values.filter(_ != null)))
).flatten.reduceLeftOption(FilterApi.or)
} else {
makeInPredicate.lift(fieldType).map(_(fieldNames, values, statistics))
makeInPredicate.lift(fieldType).map(_(fieldNames, values))
}
} else {
None
Expand Down
Loading