[
https://issues.apache.org/jira/browse/SPARK-19629?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Hyukjin Kwon resolved SPARK-19629.
----------------------------------
Resolution: Not A Problem
{quote}
What could other solutions be, if this is not a bug?
{quote}
Setting {{openCostInBytes}} above might be a workaround for it. and i don't
think it is a bug and they are guaranteed to be same.
I am resolving this JIRA. please reopen this if I am mistaken.
> Partitioning of Parquet is not considered correctly at loading in local[X]
> mode
> -------------------------------------------------------------------------------
>
> Key: SPARK-19629
> URL: https://issues.apache.org/jira/browse/SPARK-19629
> Project: Spark
> Issue Type: Bug
> Components: Input/Output, Spark Core
> Affects Versions: 2.0.0, 2.1.0
> Environment: Tested using docker run
> gettyimages/spark:1.6.1-hadoop-2.6 and
> docker run gettyimages/spark:2.1.0-hadoop-2.7.
> Reporter: Navige
> Priority: Minor
>
> Running the following two examples will lead to different results depending
> on whether the code is run using Spark 1.6 or Spark 2.1.
> h1.What does the example do?
> - The code creates an exemplary dataframe with random data.
> - The dataframe is repartitioned and stored to disk.
> - Then the dataframe is re-read from disk.
> - The number of partitions of the dataframe is considered.
> h1. What is the/my expected behaviour?
> The number of partitions specified when storing the dataframe should be the
> same as when re-loading the dataframe from disk.
> h1. Differences in Spark 1.6 and Spark 2
> On Spark 1.6 the partitioning is kept, i.e., the code example will return 10
> partitions as specified using npartitions; on Spark 2.1 the number of
> partitions will equal the number of local nodes specified when starting Spark
> (using local[X] as master).
> Looking at the data produced, in both Spark versions the number of files in
> the parquet directory is the same - so Spark 2 produces so many files as the
> number partitions when storing, but when reading in Spark 2, the number of
> partitions is messed up.
> h1.Minimal code example
> {code:none}
> # run on Spark 1.6
> import scala.util.Random
> import org.apache.spark.sql.types.{StructField, StructType, FloatType}
> import org.apache.spark.sql.Row
> val rdd = sc.parallelize(Seq.fill(100)(Row(Seq(Random.nextFloat()): _*)))
> val df = sqlContext.createDataFrame(rdd, StructType(Seq(StructField("test",
> FloatType))))
> val npartitions = 10
> df.repartition(npartitions).write.parquet("/tmp/test1")
> val read = sqlContext.read.parquet("/tmp/test1")
> assert(npartitions == read.rdd.getNumPartitions) //true on Spark 1.6
> {code}
> {code:none}
> # run on Spark 2.1
> import scala.util.Random
> import org.apache.spark.sql.types.{StructField, StructType, FloatType}
> import org.apache.spark.sql.Row
> val rdd = sc.parallelize(Seq.fill(100)(Row(Seq(Random.nextFloat()): _*)))
> val df = spark.sqlContext.createDataFrame(rdd,
> StructType(Seq(StructField("test", FloatType))))
> val npartitions = 10
> df.repartition(npartitions).write.parquet("/tmp/test1")
> val read = spark.sqlContext.read.parquet("/tmp/test1")
> assert(npartitions == read.rdd.getNumPartitions) //false on Spark 2.1
> {code}
> h1.What could other solutions be, if this is not a bug?
> If this is intended, what about introducing a parameter at reading time,
> which specifies whether the data should truly be repartitioned (depending on
> the number of nodes) or should be read "as-is".
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
