When IndexWriter.optimize() is called, ConcurrentMergeScheduler will
run the requested merges with background threads and optimize() will
wait for these merges to complete.
If a merge hits an exception, it records the root cause exception such
that optimize can then retrieve this root cause and throw its own
exception, with the root cause.
But there is a bug: sometimes, the fact that an exception occurred on
a merge is recorded, but the root cause is missing. In this cause,
optimize() still throws an exception (correctly indicating that the
optimize() has not finished successfully), but it's not helpful
because it's missing the root cause. You must then go find the root
cause in the JRE's stderr logs.
This has hit a few users on this lists, most recently:
I found the isssue, and finally got a unit test to intermittently show
it. It's a simple thread safety issue: in a finally clause in
IndexWriter.merge we record the fact that the merge hit an exception
before actually setting the root cause, and then only in
ConcurrentMergeScheduler's exception handler do we set the root
cause. If the optimize thread is scheduled in between these two, it
can throw an exception missing its root cause.
The fix is straightforward. I plan to commit to 2.4 & 2.9.