We were seeing some weird behaviour with limit based scan queries. In particular, we see the following:
In the table uuid is partition key, and it has a clustering key as well.
So the uuid "c348aaec2f1e4b85" should be the second one in the limit query. After some investigation, it seems to me like the issue is in the way DataLimits handles static rows. Here is a patch for trunk (https://github.com/sharvanath/cassandra/commit/9a460d40e55bd7e3604d987ed4df5c8c2e03ffdc) which seems to fix it for me. Please take a look, seems like a pretty critical issue to me.
I have forked the dtests for it as well. However, since trunk has some failures already, I'm not fully sure how to infer the results.