[GitHub] [accumulo] ctubbsii commented on issue #3596: Decide what to do with property tserver.last.location.mode

["ctubbsii (via GitHub)" <gi...@apache.org>]

ctubbsii commented on issue #3596:
URL: https://github.com/apache/accumulo/issues/3596#issuecomment-1661788938

   While it may still have an effect, and could still be useful, I would like to point out that the reason the property exists was mainly to change the default assignment behavior to the `assignment` mode. Prior to this property existing, the default behavior was hard-coded in the `compaction` mode. The compaction mode itself was a best-effort to achieve data locality. It's not clear that the benefits were very substantial, though. Tablets can migrate around, and the location where data was most recently compacted may not even be the location where most of the tablet's data is stored. It's also possible that none of the user's data is local, even with the mode set to compaction... as that's dependent on very specific assumptions about HDFS behavior that may not hold (the assumption that there's a local copy made for each block replica). A lot of those assumptions fall through when you consider erasure coding instead of replication, data S3 instead of HDFS, differences in user configu
 ration for replication, the fact that a data node might not even be running on the same node as the tserver, etc.
   
   I think this property could be removed if users have more metrics about where tablet files are being compacted, and could use that information to exercise fine-grained control over tablet assignments by some other means. Then again, in a world where most of Accumulo is offline and hosted only ondemand, I'm not sure it's really going to matter, because we're no longer going to care as much about balancing thrashing or churn on assignment locations due to rebooting a cluster.


-- 
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: notifications-unsubscribe@accumulo.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org