We have some data that we are exporting from our HDFS cluster to S3 with some help from Spark. The final RDD command we run is:
csvData.saveAsTextFile("s3n://data/mess/2014/11/dump-oct-30-to-nov-5-gzip", classOf[GzipCodec]) We have our 'spark.local.dir' set to our large ephemeral partition on each slave (on EC2), but with compression on an intermediate format seems to be written to /tmp/hadoop-root/s3. Is this a bug in Spark or are we missing a configuration property? It's a problem for us because the root disks on EC2 xls are small (~ 5GB).