There exists a defect that if a recovering JVM does the following steps:

doRecoveryScan()
waitForOrphanInterval()
doRecoveryScan()

Before XATerminator::recover is called in the root coordinator then bottom-up recovery will be told there is NoTransaction which is infers as a rollback.

I saw two options, one was to have TransactionCacheItem try to peek in the objectstore for ServerTransaction/JCA classes. The alternative is to add a new recovery module that can load the ServerTransaction/JCA classes periodically so the transaction won't be lost.