Exactly. Restrictive query model and data rw patterns required all tech decisions to be bounded by couch requirements (which is a no no normally) but in 2 years, 40TB and going from 1 to 14 nodes seamlessly, there was no lost data or crippling cluster events.

Thats a big deal.

"NoSQL" is meaningless. Instead, look at programming language terminology: general-purpose languages vs. domain-specific languages.

CouchDB is a domain-specific database.

I prefer to think of it as a kernel, especially in the context of all the systems built in top of it. Not just Couchbase and Cloudant but also various users who have custom wrappers (Meebo, Bing back in the day, new stuff like Refuge and mobile versions...)

This is a good way to look at it, I think. People often ask if I'd use CouchDB again if I was building another similar system - I would, but I'd restrict it to a subset of the problem that was best-suited to CouchDB's strengths/quirks.

(note: I gave the linked talk)

and what would that subset be in your opinion?