For some high capacity ingest, the desired path is to do some pre-splits, bulk import, and then let it naturally split down the rest of the way. If all of the pre-split tablets made split evenly, then the system will have continuous ranges bundled together on tservers. This poor distribution can impact performance depending on the operations performed. This could be handled in the load balancer, but it could be tricky. You don't want to randomly reassign tablets with any sort of frequency. Rather, you want to do a one-time operation in doing so. Given the initial assignment code is a bit random (needs to be validated), this could easily be done by offlining a table, purging all location records for it from the !METADATA table, and bringing it back online. The balancer will assign the table randomly, at which point the user could force a major compaction to restablish locality (as well some permanence in tablet assignment).