Naomi,
>
>I can eliminate transactions in my tests, but I can not eliminate them in a real process otherwise I have a risk of messing up database pretty badly.

Do so if it is less than 1H to test for certain. IF the error DOESN'T happen with transactions inactive (RI and normal code!) you have an important datum for further research.

>I may disable RI, though (right now we're running with RI enabled which uses its own transactions). I was able to reproduce error on several machines, but not every time. I would say, I have this error on each thrid try with big (more than 5K) files.

With that error rate (guesstimate .20 for single run) a errorfree runs of >15 should be "scientific". But the idea is to have the index recreated on the new test machine to eliminate copy effects of an index possibly compromised by faulty HW.

regards

thomas