Personally I'd be worried that the underlying database libraries were reentrant. Safer might be to have a single DB interaction thread pushing units of work onto a shared queue that separate worker threads then handle (your typical producer/consumers "pattern").