ok, I kind of found the magic bullet , but you can only use it to shoot your enemy close really close range :)
for read path, the thrift API already limits the output to a list of columns, so it does not make sense to use maps in the internal operations. plus the return CF on the read path is not going to be modified/shared by any other threads, so synchronization is not necessary. so the solution is to modify ColumnFamilyStore so that getTopLevelColumns takes a returnCF param, instead of always constructing it inside with ColumnFamily.create(). so only read path behavior is changed. in read path, we pass in a FastColumnFamily implementation, which uses an ArrayList internally to store sorted columns, and do binary search to insert , and merge to addAll(column). I tried out this, it's about 50% faster on rows with 3000 cols. Jonathan: do you think this is a viable approach? the only disadvantage is a slight change to getTopLevelColumns so we have 2 flavors of this method Thanks Yang On Wed, Jun 29, 2011 at 5:51 PM, Jonathan Ellis <jbel...@gmail.com> wrote: > On Tue, Jun 28, 2011 at 10:06 PM, Yang <teddyyyy...@gmail.com> wrote: > > I'm trying to see whether there are some easy magic bullets for a drop-in > > replacement for concurrentSkipListMap... > > I'm highly interested if you find one. :) > > -- > Jonathan Ellis > Project Chair, Apache Cassandra > co-founder of DataStax, the source for professional Cassandra support > http://www.datastax.com >
