You can turn off durability, but then you risk losing everything.
For features... forget about geo-spatial extensions - they can only do read-committed isolation level... but as they only allow for one SQL statement per transaction I suppose this might not be such a big concern.
If you were using this for something like a session management database, or something where durability is not so important then it's probably fine. Not sure I'd use it for anything that relies heavily on transaction management features.
I see other limitations, like it cannot support CHANGE COLUMN, and it can't do joins on more than 2 tables. Actually, it doesn't look like it can do FULL OUTER or RIGHT OUTER joins, either. :(
Of course a new product will have certain limitation which will be removed as the product matures.
MemSQL is fully durable (with option of sync and async durability, similar to sync and async replication) with its transactional log. If MemSQL process is killed it will recover all commited transactions on the next startup. Durability has very little performance cost on throughput as MemSQL uses only sequential I/O. And of course, durability can be completely turned off for maximum performance.
(For specific examples that I've run into recently: does it have geo-spatial extensions, and does it crash and burn and corrupt all its data if a malloc() fails)