From 7c33b0fd050f3d2b08c1cfd7efbff8166832c1af Mon Sep 17 00:00:00 2001
From: Josh Rosen <joshrosen@databricks.com>
Date: Fri, 2 Dec 2016 21:14:34 -0800
Subject: [PATCH] [SPARK-18362][SQL] Use TextFileFormat in implementation of
CSVFileFormat
## What changes were proposed in this pull request?
This patch significantly improves the IO / file listing performance of schema inference in Spark's built-in CSV data source.
Previously, this data source used the legacy `SparkContext.hadoopFile` and `SparkContext.hadoopRDD` methods to read files during its schema inference step, causing huge file-listing bottlenecks on the driver.
This patch refactors this logic to use Spark SQL's `text` data source to read files during this step. The text data source still performs some unnecessary file listing (since in theory we already have resolved the table prior to schema inference and therefore should be able to scan without performing _any_ extra listing), but that listing is much faster and takes place in parallel. In one production workload operating over tens of thousands of files, this change managed to reduce schema inference time from 7 minutes to 2 minutes.
A similar problem also affects the JSON file format and this patch originally fixed that as well, but I've decided to split that change into a separate patch so as not to conflict with changes in another JSON PR.
## How was this patch tested?
Existing unit tests, plus manual benchmarking on a production workload.
Author: Josh Rosen <joshrosen@databricks.com>
Closes #15813 from JoshRosen/use-text-data-source-in-csv-and-json.
---
.../datasources/csv/CSVFileFormat.scala | 60 ++++++++-----------
.../datasources/csv/CSVRelation.scala | 4 +-
2 files changed, 28 insertions(+), 36 deletions(-)
diff --git a/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/csv/CSVFileFormat.scala b/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/csv/CSVFileFormat.scala
index a3691158ee..e627f040d3 100644
--- a/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/csv/CSVFileFormat.scala
+++ b/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/csv/CSVFileFormat.scala
@@ -27,10 +27,12 @@ import org.apache.hadoop.mapreduce._
import org.apache.spark.TaskContext
import org.apache.spark.rdd.RDD
-import org.apache.spark.sql.SparkSession
+import org.apache.spark.sql.{Dataset, Encoders, SparkSession}
import org.apache.spark.sql.catalyst.InternalRow
import org.apache.spark.sql.catalyst.util.CompressionCodecs
import org.apache.spark.sql.execution.datasources._
+import org.apache.spark.sql.execution.datasources.text.TextFileFormat
+import org.apache.spark.sql.functions.{length, trim}
import org.apache.spark.sql.sources._
import org.apache.spark.sql.types._
import org.apache.spark.util.SerializableConfiguration
@@ -52,17 +54,21 @@ class CSVFileFormat extends TextBasedFileFormat with DataSourceRegister {
sparkSession: SparkSession,
options: Map[String, String],
files: Seq[FileStatus]): Option[StructType] = {
+ require(files.nonEmpty, "Cannot infer schema from an empty set of files")
val csvOptions = new CSVOptions(options)
// TODO: Move filtering.
val paths = files.filterNot(_.getPath.getName startsWith "_").map(_.getPath.toString)
- val rdd = baseRdd(sparkSession, csvOptions, paths)
- val firstLine = findFirstLine(csvOptions, rdd)
+ val lines: Dataset[String] = readText(sparkSession, csvOptions, paths)
+ val firstLine: String = findFirstLine(csvOptions, lines)
val firstRow = new CsvReader(csvOptions).parseLine(firstLine)
val caseSensitive = sparkSession.sessionState.conf.caseSensitiveAnalysis
val header = makeSafeHeader(firstRow, csvOptions, caseSensitive)
- val parsedRdd = tokenRdd(sparkSession, csvOptions, header, paths)
+ val parsedRdd: RDD[Array[String]] = CSVRelation.univocityTokenizer(
+ lines,
+ firstLine = if (csvOptions.headerFlag) firstLine else null,
+ params = csvOptions)
val schema = if (csvOptions.inferSchemaFlag) {
CSVInferSchema.infer(parsedRdd, header, csvOptions)
} else {
@@ -173,51 +179,37 @@ class CSVFileFormat extends TextBasedFileFormat with DataSourceRegister {
}
}
- private def baseRdd(
- sparkSession: SparkSession,
- options: CSVOptions,
- inputPaths: Seq[String]): RDD[String] = {
- readText(sparkSession, options, inputPaths.mkString(","))
- }
-
- private def tokenRdd(
- sparkSession: SparkSession,
- options: CSVOptions,
- header: Array[String],
- inputPaths: Seq[String]): RDD[Array[String]] = {
- val rdd = baseRdd(sparkSession, options, inputPaths)
- // Make sure firstLine is materialized before sending to executors
- val firstLine = if (options.headerFlag) findFirstLine(options, rdd) else null
- CSVRelation.univocityTokenizer(rdd, firstLine, options)
- }
-
/**
* Returns the first line of the first non-empty file in path
*/
- private def findFirstLine(options: CSVOptions, rdd: RDD[String]): String = {
+ private def findFirstLine(options: CSVOptions, lines: Dataset[String]): String = {
+ import lines.sqlContext.implicits._
+ val nonEmptyLines = lines.filter(length(trim($"value")) > 0)
if (options.isCommentSet) {
- val comment = options.comment.toString
- rdd.filter { line =>
- line.trim.nonEmpty && !line.startsWith(comment)
- }.first()
+ nonEmptyLines.filter(!$"value".startsWith(options.comment.toString)).first()
} else {
- rdd.filter { line =>
- line.trim.nonEmpty
- }.first()
+ nonEmptyLines.first()
}
}
private def readText(
sparkSession: SparkSession,
options: CSVOptions,
- location: String): RDD[String] = {
+ inputPaths: Seq[String]): Dataset[String] = {
if (Charset.forName(options.charset) == StandardCharsets.UTF_8) {
- sparkSession.sparkContext.textFile(location)
+ sparkSession.baseRelationToDataFrame(
+ DataSource.apply(
+ sparkSession,
+ paths = inputPaths,
+ className = classOf[TextFileFormat].getName
+ ).resolveRelation(checkFilesExist = false))
+ .select("value").as[String](Encoders.STRING)
} else {
val charset = options.charset
- sparkSession.sparkContext
- .hadoopFile[LongWritable, Text, TextInputFormat](location)
+ val rdd = sparkSession.sparkContext
+ .hadoopFile[LongWritable, Text, TextInputFormat](inputPaths.mkString(","))
.mapPartitions(_.map(pair => new String(pair._2.getBytes, 0, pair._2.getLength, charset)))
+ sparkSession.createDataset(rdd)(Encoders.STRING)
}
}
diff --git a/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/csv/CSVRelation.scala b/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/csv/CSVRelation.scala
index 52de11d403..e4ce7a94be 100644
--- a/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/csv/CSVRelation.scala
+++ b/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/csv/CSVRelation.scala
@@ -34,12 +34,12 @@ import org.apache.spark.sql.types._
object CSVRelation extends Logging {
def univocityTokenizer(
- file: RDD[String],
+ file: Dataset[String],
firstLine: String,
params: CSVOptions): RDD[Array[String]] = {
// If header is set, make sure firstLine is materialized before sending to executors.
val commentPrefix = params.comment.toString
- file.mapPartitions { iter =>
+ file.rdd.mapPartitions { iter =>
val parser = new CsvReader(params)
val filteredIter = iter.filter { line =>
line.trim.nonEmpty && !line.startsWith(commentPrefix)
--
GitLab