From: Wolfgang <wol...@ex...> - 2008-10-01 22:06:39
|
> We would thus need to synchronize access to cachedDocs, but this > seems much too expensive when compared to the benefits. I rather > prefer to completely clear the cached fields. See my fix in rev 8200: > > > http://exist.svn.sourceforge.net/exist/?rev=8200&view=rev According to the second stack trace you sent me, my previous fix for the concurrency issue (in rev 8200) wasn't sufficient. I could not reproduce the issue with a test, but I think we indeed need to synchronize access to the cached fields in LocationStep and some other query classes. I changed this and did run some tests to see if there's a measurable impact on query performance. So far, it all looks ok. The changes are here: http://exist.svn.sourceforge.net/exist/?rev=8209&view=rev Could you test this? As I said, I failed to create a test for the issue. I guess the timing between the concurrent threads needs to be very exact in order to provoke the bug. Wolfgang |