alamb commented on issue #1570: URL: https://github.com/apache/arrow-datafusion/issues/1570#issuecomment-1245882229
Thanks for the info @milenkovicm In general I think database systems tend to try and shy away from using mmap for spilling operations, for the reasons very clearly articulated in "Are You Sure You Want to Use MMAP in Your Database Management System?" https://db.cs.cmu.edu/mmap-cidr2022/ I think the current state of the art / best practice for dealing with grouping when the hashtable doesn't fit in memory is to: 1. Sort the data by groups and write that data to a spill file (ideally sharing / reusing the code used for externalized sort) 2. Create a new hash table, spilling repeatedly if needed 3. When all the input has been processed, then the spill files are read back and merged (again sharing code for externalized sort) and a merge group by is performed. Thanks to @yjshen we have much of the spilling / merging code already. We are lacking: - [ ] merge group by (where the groups are guaranteed to be contiguous in the input stream) - [ ] The orchestration logic to spill and then re-merge Does that make sense? I know it is a high level description -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
