According to: http://blog.erdemagaoglu.com/post/4605524309/lzo-vs-snappy-vs-lzf-vs-zlib-a-comparison-of
performance of snappy and lzf were on-par to each other. Maybe lzf has lower memory requirement. On Wed, May 18, 2016 at 7:22 AM, Serega Sheypak <serega.shey...@gmail.com> wrote: > Switching from snappy to lzf helped me: > > *spark.io.compression.codec=lzf* > > Do you know why? :) I can't find exact explanation... > > > > 2016-05-18 15:41 GMT+02:00 Ted Yu <yuzhih...@gmail.com>: > >> Please increase the number of partitions. >> >> Cheers >> >> On Wed, May 18, 2016 at 4:17 AM, Serega Sheypak <serega.shey...@gmail.com >> > wrote: >> >>> Hi, please have a look at log snippet: >>> 16/05/18 03:27:16 INFO spark.MapOutputTrackerWorker: Doing the fetch; >>> tracker endpoint = >>> NettyRpcEndpointRef(spark://mapoutputtrac...@xxx.xxx.xxx.xxx:38128) >>> 16/05/18 03:27:16 INFO spark.MapOutputTrackerWorker: Got the output >>> locations >>> 16/05/18 03:27:16 INFO storage.ShuffleBlockFetcherIterator: Getting 30 >>> non-empty blocks out of 30 blocks >>> 16/05/18 03:27:16 INFO storage.ShuffleBlockFetcherIterator: Started 30 >>> remote fetches in 3 ms >>> 16/05/18 03:27:16 INFO spark.MapOutputTrackerWorker: Don't have map >>> outputs for shuffle 1, fetching them >>> 16/05/18 03:27:16 INFO spark.MapOutputTrackerWorker: Doing the fetch; >>> tracker endpoint = >>> NettyRpcEndpointRef(spark://mapoutputtrac...@xxx.xxx.xxx.xxx:38128) >>> 16/05/18 03:27:16 INFO spark.MapOutputTrackerWorker: Got the output >>> locations >>> 16/05/18 03:27:16 INFO storage.ShuffleBlockFetcherIterator: Getting 1 >>> non-empty blocks out of 1500 blocks >>> 16/05/18 03:27:16 INFO storage.ShuffleBlockFetcherIterator: Started 1 >>> remote fetches in 1 ms >>> 16/05/18 03:27:17 ERROR executor.Executor: Managed memory leak detected; >>> size = 6685476 bytes, TID = 3405 >>> 16/05/18 03:27:17 ERROR executor.Executor: Exception in task 285.0 in >>> stage 6.0 (TID 3405) >>> >>> Is it related to https://issues.apache.org/jira/browse/SPARK-11293 >>> >>> Is there any recommended workaround? >>> >> >> >