(disclaimer: I haven't looked at Cassandra in 6 months)

Riak doesn't involve Java. It sounds trite, but that was a huge plus for our team. It has commit hooks. Search. Builtin JS VMs for mapreduce. More straightforward administration. Most operations are automatic or on-the-fly. Even though I think Cassandra+Hadoop is a technically superior product for managing absolutely huge gobs of multidimensional data (especially with its support for range queries), Riak's ease of use is pretty sweet.

Plus the team is superhumanly helpful about addressing bugs and gotchas. Everything's on Github, and contributing is pretty easy.

I can attest to their support being phenomenal. I found a critical bug last week and got a patch within hours of reporting it to their customer advocate. You don't see that kind of turnaround very often.

Riak Search is in fact great to work with.

Something comparable for Cassandra exists with Solandra (https://github.com/tjake/Lucandra/). A major difference is that you'll have to post your documents manually to the Solr Instance running with Solandra while Riaks post-receive hooks automatically do that for you.

Also: the setup is by far not as big of a pain with Riak

IANAE, but while reading over the riak docs for an upcoming project, the concept of links is one that stood out. http://blog.basho.com/2010/03/25/schema-design-in-riak---rel...

Links are awesome in a small dataset, but on a larger cluster, the probability that each link walk means both an extra seek and a network roundtrip makes them less attractive.

Local graph databases avoid the extra seeks by keeping extra copies of the destination vertex immediately next to the source vertex on disk, but I don't think Riak is doing this, or that it is actually practical in an eventually consistent system.

Some of the stuff on that page is no longer current and some of it is pretty underwhelming. I guess the re-balancing story and vector clocks sound nice?

It's a wiki, if there is information you see is incorrect, please fix it.

From my perspective, and not an expert on cassandra, but: -- Conceptually more straightforward for engineers, thus less code complexity and less bugs (and the cost of maybe less performance. I'll spend dollars on more nodes over dollars to deal with engineering complexity.) -- Operationally cassandra sounds like a PITA. I haven't run it, so I could be misinformed. -- Engineering philosophy. The Basho style is more in line with my style than the people behind cassandra. The priorities and emphasizes are different.