This is sort of a last-ditch effort. We have Oracle SOA on Fusion, 220.127.116.11.0 (+ patches on a separate server) running, and nearing the point where we're taking it into production. I've posted on OTN and talked to numerous folks at Oracle about this problem, but no one can seem to help me.
When under production-level load (simulating about 1 request every 30 seconds) Weblogic encounters STUCK threads and SOA simply stops processing composite instances. The composite that fails changes from time to time, and the logs indicate that deadlocked thread is trying to load a WSDL or referenced schema file in MDS. We have several shared schema and WSDL files across many composites on multiple partitions. Each SOA partition has its own MDS repository. The end result is that the Weblogic server needs to be restarted before processing can proceed. There's nothing logged in the audit trail and nothing blatantly evident in the server's diagnostics logs. I upped the logging to the most verbose level, and can't personally see a pattern. I have since given Oracle the best possible logs and thread dumps available.
In the audit trail, I can click into the faulting mediator/composite for details, and the call graph shows that it calls the related service, but there's no link, and viewing the invoked composite in EM does not show an entry related to that invocation.
Is this a familiar symptom to anyone?