Disable parquet metadata summary in Spark

9
votes

I have a spark job (for 1.4.1) receiving a stream of kafka events. I would like to save them continuously as parquet on tachyon.

val lines = KafkaUtils.createStream(ssc, zkQuorum, group, topicMap).map(_._2)

lines.window(Seconds(1), Seconds(1)).foreachRDD { (rdd, time) =>
  if (rdd.count() > 0) {
    val mil = time.floor(Duration(86400000)).milliseconds
    hiveContext.read.json(rdd).toDF().write.mode(SaveMode.Append).parquet(s"tachyon://192.168.1.12:19998/persisted5$mil")
    hiveContext.sql(s"CREATE TABLE IF NOT EXISTS persisted5$mil USING org.apache.spark.sql.parquet OPTIONS ( path 'tachyon://192.168.1.12:19998/persisted5$mil')")
  }
}

however I see that as time goes on, on every parquet write, spark goes through each 1 sec parquet parts, which get slower and slower

15/08/22 22:04:05 INFO : open(tachyon://192.168.1.12:19998/persisted51440201600000/part-r-00000-db03b24d-6f98-4b5d-bb40-530f35b82633.gz.parquet, 65536)
15/08/22 22:04:05 INFO : open(tachyon://192.168.1.12:19998/persisted51440201600000/part-r-00000-3a7857e2-0435-4ee0-ab2c-6d40224f8842.gz.parquet, 65536)
15/08/22 22:04:05 INFO : open(tachyon://192.168.1.12:19998/persisted51440201600000/part-r-00000-47ff2ac1-da00-4473-b3f7-52640014bc5b.gz.parquet, 65536)
15/08/22 22:04:05 INFO : open(tachyon://192.168.1.12:19998/persisted51440201600000/part-r-00000-61625436-7353-4b1e-bb8d-e8afad3a582e.gz.parquet, 65536)
15/08/22 22:04:05 INFO : open(tachyon://192.168.1.12:19998/persisted51440201600000/part-r-00000-e711aa9a-9bf5-41d5-8523-f5edafa69626.gz.parquet, 65536)
15/08/22 22:04:05 INFO : open(tachyon://192.168.1.12:19998/persisted51440201600000/part-r-00000-4e0cca38-cf75-4771-8965-20a30c863100.gz.parquet, 65536)
15/08/22 22:04:05 INFO : open(tachyon://192.168.1.12:19998/persisted51440201600000/part-r-00000-d1510ed4-2c99-43e2-b3d1-38d3d54e626d.gz.parquet, 65536)
15/08/22 22:04:05 INFO : open(tachyon://192.168.1.12:19998/persisted51440201600000/part-r-00000-022d1918-392d-433f-a7f4-074e46b4460f.gz.parquet, 65536)
15/08/22 22:04:05 INFO : open(tachyon://192.168.1.12:19998/persisted51440201600000/part-r-00000-cf71f5d2-ba0e-4729-9aa1-41dad5d1d08f.gz.parquet, 65536)
15/08/22 22:04:05 INFO : open(tachyon://192.168.1.12:19998/persisted51440201600000/part-r-00000-ce990b1e-82cc-4feb-a162-ac3ddc275609.gz.parquet, 65536)

I came to the conclusion that this is due to the update of summary data, I believe spark do not make use of them. so I would like to disable it

parquet sources shows that I should be able to set "parquet.enable.summary-metadata" to false.

now, I have tried setting it like this, right after creating hiveContext

hiveContext.sparkContext.hadoopConfiguration.setBoolean("parquet.enable.summary-metadata", false)
hiveContext.sparkContext.hadoopConfiguration.setInt("parquet.metadata.read.parallelism", 10)

but without success, I also still get logs showing a parallelism of 5 (default).

What is the correct way to disable summary data in spark with parquet?

2
apache-sparkparquet

2 Answers

14
votes

setting "parquet.enable.summary-metadata" as text ("false" and not false) seems to work for us.

By the way Spark does use the _common_metadata file (we copy that over manually for repetitive jobs)

14
votes

Spark 2.0 doesn't save metadata summaries by default any more, see SPARK-15719.

If you are working with data hosted in S3, you may still find parquet performance hit by parquet itself trying to scan the tail of all objects to check their schemas. That can be disabled explicitly

sparkConf.set("spark.sql.parquet.mergeSchema", "false")