You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@cassandra.apache.org by Brayton Thompson <th...@grnoc.iu.edu> on 2010/11/18 21:50:52 UTC
Load Vs Disk Space Usage
We're playing around with Cassandra trying to get a feel for it. Can someone please explain the difference between load (from nodetool) and whats actually stored on disk? Sometimes these number mirror each other and sometimes the disk usage is up to 2x the load reported. as you can see below...
Cassandra 1 Keyspace Data Dir. Size (MB) System Data Dir Size (MB) Commit Log Data Dir Size (MB) Load (reported by ring) (MB)
Initial 0 0.11 0.03 0.04
Run 1 254 0.11 344 253.23
Run 2 634 0.11 302 632.84
Run 3 1500 0.11 260 1012.46
Run 4 1400 0.11 219 1360
Run 5 2200 0.11 305 1610
Run 6 4000 0.11 264 1980
Run 7 2400 0.11 350 2350
Run 8 2800 0.11 309 2720
Run 9 3600 0.11 267 3090
Run 10 3400 0.11 354 3340
Run 11 3800 0.11 312 3710
Run 12 6100 0.11 271 4080
Run 3,5,6,9, and 12 don't seem to match up well. Can someone explain this please?
Thank you.
Re: Load Vs Disk Space Usage
Posted by Peter Schuller <pe...@infidyne.com>.
> We're playing around with Cassandra trying to get a feel for it. Can someone please explain the difference between load (from nodetool) and whats actually stored on disk? Sometimes these number mirror each other and sometimes the disk usage is up to 2x the load reported. as you can see below...
[snip]
> Run 3,5,6,9, and 12 don't seem to match up well. Can someone explain this please?
Probably there are obsolete sstables that have not yet been removed.
Removal of sstables is somewhat delayed because it relies on GC to
avoid synchronization complexities in the implementation. See:
http://wiki.apache.org/cassandra/MemtableSSTable
I believe sstables that are obsolete will not count towards load.
You can either trigger the GC, restart the cassandra nodes, or just
wait until they disappear (generating enough activity to trigger a CMS
sweep of the heap should be enough, assuming you use CMS).
--
/ Peter Schuller