Thanks a lot. It has became clear for me. From iPhone
On 2011/04/06, at 23:51, Sylvain Lebresne <sylv...@datastax.com> wrote: > On Tue, Apr 5, 2011 at 9:03 PM, Maki Watanabe <watanabe.m...@gmail.com> wrote: >> Thanks Sylvain, it's very clear. >> But should I still need to force major compaction regularly to clear >> tombstones? >> I know that minor compaction clear the tombstones after 0.7, but >> maximumCompactionThreshold limits the maximum number of sstable which >> will be merged at once, so to GC all tombstones in all sstable in >> gc_grace_period, it is safe to run "nodetool compact" at least once in >> gc_grace_period, isn't it? > > You don't *need* tombstones to be cleared within gc_grace_period. What you > need is to make sure for a given tombstone t, that each node will get t within > gc_grace_period. This means that if a node dies, you need it to be up again > and have nodetool repair ran before gc_grace_period, otherwise there may > be some tombstones that this node will never see (and thus deleted data > could be resurrected by this node). > > So repair should be run at least once in gc_grace_period to be on the safe > side. > Compact is not necessary however. The only downside of not running compact > regularly is that some tombstones may take longer to be removed (since minor > compaction are potentially less efficient at removing them), which > really only impact > disk space usage. And given that major compaction are fairly heavy on > ressource > usage and have that nasty effect of producing only one huge sstable, my advice > would be to not run major compaction unless you have good reason to suspect > you need it. > > -- > Sylvain > >> >> maki >> >> 2011/4/6 Sylvain Lebresne <sylv...@datastax.com>: >>> On Tue, Apr 5, 2011 at 12:01 AM, Maki Watanabe <watanabe.m...@gmail.com> >>> wrote: >>>> Hello, >>>> On reading O'Reilly's Cassandra book and wiki, I'm a bit confusing on >>>> nodetool repair and compact. >>>> I believe we need to run nodetool repair regularly, and it synchronize >>>> all replica nodes at the end. >>>> According to the documents the "repair" invokes major compaction also >>>> (as side effect?). >>> >>> Those documents are wrong then. A repair does not trigger a major >>> compaction. The only thing that makes it similar to a major compaction is >>> that it will iterate over all the sstables. But for instance, you won't end >>> up with one big sstable at the end of repair as you would with a major >>> compaction. >>> >>>> Will this "major compaction" apply on replica nodes too? >>>> >>>> If I have 3 node ring and CF of RF=3, what should I do periodically on >>>> this system is: >>>> - nodetool repair on one of the nodes >>>> or >>>> - nodetool repair on one of the nodes, and nodetool compact on 2 of the >>>> nodes >>>> ? >>> >>> So as said, repair and compact are independent. You should >>> periodically run nodetool >>> repair (on one of your nodes in your case as you said). However, it is >>> not advised anymore >>> to run nodetool compact regularly unless you have a good reason to. >>> >>> -- >>> Sylvain >>> >>
