It should be done automatically, unless you're running a custom load
balancer that balances tablets different. However, there are some bugs in a
few releases of 1.4 which broke that. A restart of the master process
should be all you need to get things balancing again.
Okay, thank you! Is there a means of checking the load of a given, if not all, tablet(s)? I have an uneven distribution over my tablets where a few have hundred GBs and the rest only have a few hundred MBs leading me to believe that some of the tablets are being under-used, hence the question about re-balancing.
Wait, are you trying to balance out your tablets over your tablet servers
or your data over your tablets?
If the former, my aforementioned suggestion should help. If the latter,
it's more complicated.
When you create a table, it's default one tablet, or more if you supply
split points. As you ingest, once the size on disk of a tablet exceeds the
split threshold, it divides in half as evenly as it can. It will not divide
mid-row though. So if you have a few giant tablets and a lot of tiny
tablets, this is most likely due to high cardinality rows. The only real
thing you can deal with there is to change your key format and reingest.
> Okay, thank you! Is there a means of checking the load of a given, if not
> all, tablet(s)? I have an uneven distribution over my tablets where a few
> have hundred GBs and the rest only have a few hundred MBs leading me to
> believe that some of the tablets are being under-used, hence the question
> about re-balancing.
> Thanks again,
> View this message in context:
> http://apache-accumulo.1065345.n5.nabble.com/Rebalance-table-over-all-tablet-servers-tp5393p5396.html > Sent from the Developers mailing list archive at Nabble.com.
I'm trying to do the latter. Is there any way to check the load of a tablet? Or I guess how close a tablet is to the defined/default split? How much of the tablet space is used? I've looked into the "du" command, but that only gives the size of a table and doesn't talk about how it is split up.