> It's also quite possible that the mis-handling of exceptions caused the
> problem by leading to the acceptance of commiting of a partially complete
> operator.

In theory that should never happen. Personally I have suspect that there is
some issues with transaction handling, namely that somehow distributed
transactions are used and part of the deletes are executed in one
transaction branch and other in another one.

But the best would be if you can create a test case to reproduce the issue.

Roman