Details
-
Improvement
-
Status: Open
-
Normal
-
Resolution: Unresolved
-
None
-
None
Description
At the SF Summit this year, Sylvain suggested that C* drops a very large amount of write throughput on the floor for multi-partition mutations because they are broken into RowMutations and executed individually. Stress tests that I've run show 10X the throughput for 1-row x 1000-col writes versus 1000-row x 1-col writes. We have a core high-write-skew use case which involves fan-out-on-write against hundreds or up to thousands of keys at a time currently implemented in Redis as it doesn't seem to suffer from the issue. Would love to be able to move this to C* at some point.
This is likely a pretty large undertaking as it would require touching a large portion of the write path, but I figure I'd put it here for comment and/or debate at this point.