I am having some problems getting to JIRA, will update the issue when I get online.
I now have an environment where I can reproduce DERBY-5780. I am actively working on this one, and think I understand the problem.
The commits are working fine, but we destroy/abort on each identity column insert and that abort is sync writing where it did not use to due to the change for sequences. I can either make there be an abortNoSync(), or change transactions to indicate on creation whether the default behavior is to sync or not. I think the second option is the right way to go but may involve some code rototil, so may have a fix this week, but more likely a tested fix to trunk by next tuesday given the holiday. On 5/23/2012 3:52 PM, Rick Hillegas wrote:
I have created a 10.9 branch and built a release candidate. However, I do not expect to publish this candidate. Some late-breaking performance measurements show a large, unexpected throughput drop for concurrent users of identity columns. See DERBY-5471 and DERBY-5780. It is likely that we will want to address this regression before producing a new release. The fix, of course, will have to be ported to 10.9 now that the branch has been created. Please stand by... Thanks, -Rick
