See my post about the projects: they are very young, just little over half-year old - performance was not focus yet. I would definitely have them to be able to handle more data more efficiently, however, goal more on simplicity of use than on ability to process really huge amounts of data (like telco data - background where I come from).

Before I answer your questions (I assume that you are referring to Cubes - OLAP framework), I think it would be good to note, that Cubes has pluggable backends. Currently simple denormalisation-based SQL backend and MongoDB backend are implemented. I want to have them more advanced.

* Will data be fetched from the database for each query?

- currently yes, however we did some experiments with plain HTTP caching of Cubes/Slicer server and it worked pretty nicely for our current needs

* Is it possible to have dimensions with millions of values and expect reasonable query times?

- not tested yet

* Looks like it supports advanced topologies and hierarchies. How will dimensions with a high carnality affect performance?

- right, it supports hierarchies, however same as above: not tested yet for performance

I am open to any commeents/suggestions regarding the framework(s).

Stefan Urbanek, @Stiivi on Twitter (author of Cubes)