Also, there are plenty of
compactions running - it just seems like the number of pending tasks
is never affected.
Is there ever a time when the pending count is non zero but nodetool compactionstats does not show any running tasks ? 

If compaction cannot keep up you may be generating data faster than LCS can compact it. What sort of disks do you have?

What is the min_sstable_size ? The old default was 5, the new one is 130. The higher value will result in less IO. 

Cheers

-----------------
Aaron Morton
New Zealand
@aaronmorton

Co-Founder & Principal Consultant
Apache Cassandra Consulting
http://www.thelastpickle.com

On 24/10/2013, at 3:16 AM, Russ Garrett <russ@garrett.co.uk> wrote:

On 23 October 2013 15:05, Michael Theroux <mtheroux2@yahoo.com> wrote:
When we made a similar move, for an unknown reason (I didn't hear any feedback from the list when I asked why this might be), compaction didn't start after we moved from SizedTiered to leveled compaction until I ran "nodetool compact <keyspace> <column-family-converted-to-lcs>".

Yup, I understand that's this issue (the fix is still unreleased):
https://issues.apache.org/jira/browse/CASSANDRA-6092

However, my suspicion is that adding/removing nodes has probably
kickstarted the compaction anyway. I've tried issuing "nodetool
compact" with no obvious improvement. Also, there are plenty of
compactions running - it just seems like the number of pending tasks
is never affected.

Good point on the version. I'm on 2.0.1.

--
Russ Garrett
russ@garrett.co.uk