I have a simple ETL Spark job running on AWS EMR with Spark 2.2.1 . The input data is HBase files in AWS S3 using EMRFS, but there is no HBase running on the Spark cluster itself. It is restoring the HBase snapshot into files on disk in another S3 folder used for temporary storage, then creating an RDD over those files using HBase's TableSnapsotInputFormat class. There is a large number of HBase regions, around 12000, and each region gets translated to one Spark task/partition. We are running in YARN mode, with one core per executor, so on our 120 node cluster we have around 1680 executors running (not the full 1960 as YARN only gives us so many containers due to memory limits).
This is a simple ETL job that transforms the HBase data into Avro/Parquet and writes to disk, there are no reduces or joins of any kind. The output Parquet data is using Snappy compression, the total output is around 7 TB while we have about 28 TB total disk provisioned in the cluster. The Spark UI shows no disk storage being used for cached data, and not much heap being used for caching either, which makes sense because in this simple job we have no need to do RDD.cache as the RDD is not reused at all.
So lately the job has started failing because close to finishing, some of the YARN nodes start running low on disk and YARN marks them as unhealthy, then kills all the executors on that node. But the problem just moves to another node where the tasks are relaunched for another attempt until after 4 failures for a given task the whole job fails.
So I am trying to understand where all this disk usage is coming from? I can see in Ganglia that disk is running low the longer the job runs no matter which node I look at. Like I said the total output size of the final output in hdfs is only around 7 TB while we have around 28 TB of disk provisioned for hdfs.
Any advice or pointers for where to look for the large disk usage would be most appreciated.