Recently, the community has actively been working on this. The JIRA to follow is: https://issues.apache.org/jira/browse/SPARK-25299. A group of various companies including Bloomberg and Palantir are in the works of a WIP solution that implements a varied version of Option #5 (which is elaborated upon in the google doc linked in the JIRA summary).
On Wed, Dec 19, 2018 at 5:20 AM <marek-simu...@seznam.cz> wrote: > Hi everyone, > we are facing same problems as Facebook had, where shuffle service is > a bottleneck. For now we solved that with large task size (2g) to reduce > shuffle I/O. > > I saw very nice presentation from Brian Cho on Optimizing shuffle I/O at > large scale[1]. It is a implementation of white paper[2]. > Brian Cho at the end of the lecture kindly mentioned about plans to > contribute it back to Spark[3]. I checked mailing list and spark JIRA and > didn't find any ticket on this topic. > > Please, does anyone has a contact on someone from Facebook who could know > more about this? Or are there some plans to bring similar optimization to > Spark? > > [1] https://databricks.com/session/sos-optimizing-shuffle-i-o > [2] https://haoyuzhang.org/publications/riffle-eurosys18.pdf > [3] > https://image.slidesharecdn.com/5brianchoerginseyfe-180613004126/95/sos-optimizing-shuffle-io-with-brian-cho-and-ergin-seyfe-30-638.jpg?cb=1528850545 >