On Thu, Feb 28, 2013 at 10:03 AM, Denis <de...@camfex.cz> wrote: > Hi, Keith > > Thanks for the hint to use iterator for that. > > Although this approach can help in my case, it may be tricky to use it > in other cases, because (as far as I understood), with iterator one > can throttle the uncompressed data and actual throttling of the hard > disk operations would depend on the data compressibility in the tablet > being compacted and can vary widely.
Thats a good point. I posted this on the ticket to move the discussion there. > > I opened the ticket ACCUMULO-1128. > > On 2/28/13, Keith Turner <ke...@deenlo.com> wrote: >> Can you open a ticket for this? We could post the iterator as a patch >> to that ticket. Then people could use that patch. >> >> On Thu, Feb 28, 2013 at 8:27 AM, Denis <de...@camfex.cz> wrote: >>> Hi. >>> >>> Major compaction loads hard disks very hard, even with >>> tserver.compaction.major.concurrent.max=1 >>> >>> Besides nice peaks on the Load Average and IOstat graps during major >>> compaction, such high load also badly affects query performance and >>> latency. >>> >>> HBase and Cassandra both have settings to throttle major compaction >>> (Cassandra's default is 16 Mb/s which is about 1/3 of the peek values >>> I see during Accumulo's major compaction) >>> >>> But it looks like Accumulo does not have such setting and I have not >>> found relevant questions in the bugtracker and maillist archive. >>> >>> How do you deal with the major compaction and its performance effects? >>