Hunchentoot is a bit odd, but works well. In particular, the fact that handlers (think, like prefix-matching dispatch functions) are global, rather than per-listener. There is, in general, an apparent tendency to use
var globals for expediency's sake, but they tend to infiltrate code, and make things harder to debug.
I concur with other commentators that PostgreSQL is worth looking at. I wrote the postgres bindings for what is now CL-SQL many years ago (hi stringbot!), and postgres is fast and reliable (aside from SBCL having interrupt hander sensitivities during ffi calls), though again, the CL-SQL interface is littered with globals, the connection pooling is a bit .. unusual .. and the reader macro syntax probably a pretty bad idea, all in all. If hunchentoot and cl-sql were mildly rewritten with an eye towards looking more like, for example, Java's JDBC and Servlet interfaces, you would have a simple base environment where you could reap most of the big CL wins immediately (DSLs, CLOS, coding closer to the speed of thought, etc.)
The 'IMHO' sources are kicking around here: they suffered a bit from WebObjects envy (hey, it was the 90's..) and do somewhat suboptimal handling of session state. It's still not clear to me how OO you really want to get with the server-side representation of a page, especially in light of all the nice JS client-side component kits that are around.
Mull.