Richard, The same problem is with sort.
I have enough disk space and tmp folder. The errors in logs tell out of memory. I wonder what does it hold in memory? Alexander From: Richard Marscher [mailto:rmarsc...@localytics.com] Sent: Tuesday, April 28, 2015 7:34 AM To: Ulanov, Alexander Cc: user@spark.apache.org Subject: Re: Scalability of group by Hi, I can offer a few ideas to investigate in regards to your issue here. I've run into resource issues doing shuffle operations with a much smaller dataset than 2B. The data is going to be saved to disk by the BlockManager as part of the shuffle and then redistributed across the cluster as relevant to the group by. So the data is going to be replicated during the operation. I might suggest trying to allocate more memory for your executors in your cluster. You might also want to look into configuring more explicitly the shuffle functionality (https://spark.apache.org/docs/latest/configuration.html#shuffle-behavior). Check the disk usage on the worker nodes, in our case we actually had small disk space to start with and were running out of temporary space for the shuffle operation. I believe you should also be able to find more clear errors in logs from the worker nodes if you haven't checked yet. On Mon, Apr 27, 2015 at 10:02 PM, Ulanov, Alexander <alexander.ula...@hp.com<mailto:alexander.ula...@hp.com>> wrote: It works on a smaller dataset of 100 rows. Probably I could find the size when it fails using binary search. However, it would not help me because I need to work with 2B rows. From: ayan guha [mailto:guha.a...@gmail.com<mailto:guha.a...@gmail.com>] Sent: Monday, April 27, 2015 6:58 PM To: Ulanov, Alexander Cc: user@spark.apache.org<mailto:user@spark.apache.org> Subject: Re: Scalability of group by Hi Can you test on a smaller dataset to identify if it is cluster issue or scaling issue in spark On 28 Apr 2015 11:30, "Ulanov, Alexander" <alexander.ula...@hp.com<mailto:alexander.ula...@hp.com>> wrote: Hi, I am running a group by on a dataset of 2B of RDD[Row [id, time, value]] in Spark 1.3 as follows: “select id, time, first(value) from data group by id, time” My cluster is 8 nodes with 16GB RAM and one worker per node. Each executor is allocated with 5GB of memory. However, all executors are being lost during the query execution and I get “ExecutorLostFailure”. Could you suggest what might be the reason for it? Could it be that “group by” is implemented as RDD.groupBy so it holds the group by result in memory? What is the workaround? Best regards, Alexander
