well, matter for what? For building software that seems to work, sure, no. But for being able to speak to its capabilities -- or even know what they are, especially in the case of distributed systems it's definitely important.

this is why so many of the new DB offerings have docs that are either misleading or wrong -- often the doc writers or even the devs themselves don't know what invariants their system can guarantee and assume several that it can't.

Depends on the problem you want to solve. CouchDB is a very good document database when you have more read than write and when writer are mostly writing on different documents. If I understand correctly replication conflict resolution is up to the application. So knowing about consensus algo is unnecessary for CouchDB because he doesn't support it.

I think it's smart to know a little about the literature; enough to know what battle you don't want to fight. It is not smart to try to develop an engine on the basis that it will need to beat the law of thermodynamic to be possible.

But also, “They did not know it was impossible so they did it”, Mark Twain

Kyle's point there is that it does matter, in that CouchDB can't achieve consensus, and would have failed Jepsen immediately.