abstraction is "the right thing to do" (tm).
Why not? The free versions exceed postgresql. If you succeed and face heavy load, then you just throw some money at the problem rather than blow hundreds of man-hours fidling with the hacks people use to scale the free DBMSes.
See http://www.commandprompt.com/support/ for an example, or http://pgexperts.com/services.html
This goes double if you are considering commercial products.
For various reasons we are forced to use $Vendor, but we are not a commercial shop and we always have the fear factor that $Vendor will price us out (they occasionally try it on by "forgetting" our academic discount). We hid all the DB interactions in two layers (one to hide the relational nature of the database, and the other to hide the specific implementation). As a result, I am in the position to tell $Vendor "How much? Well, you can give us this price or we can move to postgress - either is fine by me". And mean it. And they can tell.
I'm pretty skeptical of this approach. Cross platform development is a pain in the ass, whether we're talking databases or operating systems. I say pick the most powerful database you can and marry it. Use its features to the fullest and ignore portability. I suggest db2 or oracle.