Your first paragraph is, bluntly, incorrect. Cassandra guarantees that data will always become consistent. This is automatic [1] for normal operation, including in the face of temporary failures. Permanent failures require running a "repair" process to rebuild the failed machine from other replicas [2].
I think you've also misunderstood how quorum works; it is a quorum of the replica count, which tends to stay constant over cluster lifetime, not machine count.
You are right that the current counters are an afterthought. I linked in my concluding paragraph, where I talk about improvements for Cassandra, "A new design for distributed counters." [3]
[1] http://www.datastax.com/dev/blog/modern-hinted-handoff [2] http://www.datastax.com/docs/1.2/operations/node_repair [3] https://issues.apache.org/jira/browse/CASSANDRA-4775
The consistency claims are overblown.
Given that nodes cannot disagree on the set of replicas for a row, quorums must (by definition, and the pigeon hole principal) overlap by at least one replica.
> There's not really any way that nodes can disagree on this.
Atomic, dynamic updates to a ring are a difficult thing to do. If I'm missing some additional mechanism that ensures agreement on ring membership, please explain.
HBase for example is always consistent, but if you lose a node unexpectedly some data is not available for some time.
Have you actually experienced this issue or just making it up ?
Imagine two nodes, both initiating updates to the same row. One has observed a number of node additions and deletions. The other has not. They will perform their updates on disjoint sets.
Additions to the ring are not atomic, therefore quorums aren't.
Don't expect to run a 3-node Cassandra cluster and get much out of it in terms of availability, in the way you might run a master/slave failover setup. It's somewhat obvious, but your Cassandra deployment can't just start with a couple of nodes and scale up as you run into bottlenecks. The number of nodes needed starts to add up quickly with a replication factor of > 1 and quorum reads. And while you might say "I'm ok with eventual consistency, let's just read from a single node," if you're not reading from multiple nodes, the data may never become consistent, from what I can tell.
And counters should be marked with a big warning "not for production use". Their performance isn't great, and it nosedives as the dataset grows. (each counter update involves a read + a write) Having a node reboot can sometimes cause counters to double. They seem like basically an afterthought.