CODEWITHSUNDEEP
apache-sparkpartitioningparquet
Gaurav Shah
Gaurav Shah

Reputation: 5279

spark parquet write gets slow as partitions grow

I have a spark streaming application that writes parquet data from stream.

sqlContext.sql(
      """
        |select
        |to_date(from_utc_timestamp(from_unixtime(at), 'US/Pacific')) as event_date,
        |hour(from_utc_timestamp(from_unixtime(at), 'US/Pacific')) as event_hour,
        |*
        |from events
        | where at >= 1473667200
      """.stripMargin).coalesce(1).write.mode(SaveMode.Append).partitionBy("event_date", "event_hour","verb").parquet(Config.eventsS3Path)

this piece of code runs every hour but over time the writing to parquet has slowed down. When we started it took 15 mins to write data, now it takes 40 mins. It is taking time propotional to data existing in that path. I tried running the same application to a new location and that runs fast.

I have disabled schemaMerge and summary metadata:

sparkConf.set("spark.sql.hive.convertMetastoreParquet.mergeSchema","false")
sparkConf.set("parquet.enable.summary-metadata","false")

using spark 2.0

batch execution: empty directory enter image description here enter image description here enter image description here directory with 350 folders enter image description here enter image description here enter image description here

Upvotes: 15

Views: 17287

Answers (3)

Igor Berman
Igor Berman

Reputation: 1532

It might be due to append mode. In this mode new files should be generated with different names from already existing files, so spark lists files in s3(which is slow) every time.

We also set parquet.enable.summary-metadata a bit differently:

javaSparkContext.hadoopConfiguration().set("parquet.enable.summary-metadata", "false");

Upvotes: 0

Niros
Niros

Reputation: 642

Try to write the dataframe to EMR HDFS (hdfs://...) and then use s3-dist-cp to upload the data from HDFS to S3. Worked for me.

Upvotes: 1

Francois G
Francois G

Reputation: 11985

I've encountered this issue. The append mode is probably the culprit, in that finding the append location takes more and more time as the size of your parquet file grows.

One workaround I've found that solves this is to change the output path regularly. Merging and reordering the data from all the output dataframes is then usually not an issue.

def appendix: String = ((time.milliseconds - timeOrigin) / (3600 * 1000)).toString

df.write.mode(SaveMode.Append).format("parquet").save(s"${outputPath}-H$appendix")

Upvotes: 2

Related Questions