Dear Nick, Thanks for looking into this! I filed the ticket: https://issues.apache.org/jira/browse/HBASE-28902 and attached the jprofiler profile we used to investigate this issue.
I do not have a profile without the offending patch (I don't think i can downgrade the instance to get that). What we did now is construct our SEEK_NEXT_USING_HINT hint as follows: "bf_slicing\0" which makes it skip to the next column family due to the length now having increased. This is a bit of a hack though I think. Hope this helps, Cheers, Bram Schuur On Fri, Oct 4, 2024 at 11:16 AM Nick Dimiduk <ndimi...@apache.org> wrote: > Hi Bram, Alejandro, > > Thanks for your report. I did see your email as it came in, but I have not > taken time to refamiliarize myself with the behavior of seek hints. I’ve > approved your request for a Jira account. Please do file a ticket, link it > back to the original. If you can quantify the degredation, that would help > a lot. Can you include flame graphs of the region server with and without > the offending patch? > > Thanks, > Nick > > On Tue, 1 Oct 2024 at 14:31, Bram Schuur <bram.sch...@suse.com.invalid> > wrote: > > > Dear HBASE developers, > > > > We have a custom hbase filter that seeks (SEEK_NEXT_USING_HINT) to a next > > column family (called "cf") based on data in a cell in a prior column > > family (called "bf_slicing"). > > > > We upgraded to hbase 2.6.0 from 2.5.8, the change in this ticket > > https://issues.apache.org/jira/browse/HBASE-27788 caused a > > significant performance degradation. When comparing families here, the > 'cf' > > family is ordered lower than 'bf_slicing' due to its length, causing the > > first column family ("bf_slicing") to be fully traversed. > > > > The offending code is here: > > > > > https://github.com/apache/hbase/pull/5171/files#diff-1ec9654ed8e00f46e11430fc726f8351db59597723efa0bf1e268196f00244c6R54 > > > > The original story (HBASE_27788) mentions no seeking should be done > outside > > a column family, but our use case seems legitimate in the data model, so > we > > think this is a bug, or are we missing something? If this is a bug, I > would > > be happy to file a ticket or provide more information. > > > > Thanks in advance, > > > > Alejandro Acevedo and Bram Schuur > > >
