My personal opinion is that atm the libpq driver is already as good as or better than the pre-libpq driver. There may be a few as of yet undiscovered reliability issues or busted features, so there is some risk with the switchover. Since Dave fixed the excel bug, today I’m going to switch over my application stack to the new driver and give a more qualitative analysis.
Somewhat unintentionally I should add - it started working after what I believed were unrelated fixes :-)
Since the new driver is a much better design, (and supports the level 3 protocol!) it will only get better. I think that you can also give some really good feedback about utf-8 capabilities of the driver (expect some problems here). I hate to suggest new technology to bail you out of a nasty situation but ISTM this is the only situation with a high chance of long term success. The only other option I see is to build a stronger test suite and hang back for a bit and closely monitor progress.
I agree - the new driver does seem to be at least as good. There are some more improvements in CVS that haven't made the snapshots yet that you might want to wait for - specifically, backend error messages are reported properly now, rather than a default message being issued, a few leaks and other minor issues have been solved, and SSL support has been added (not that I expect Joel will want to use that).
How heavily have you load tested the new driver?
I've hammered it with ODBCBench, and a simple IIS app thrashed with ApacheBench. I've not made it crash yet :-) (but then I couldn't reproduce Joel's problems with the older driver).
Regards, Dave.