"Map/reduce operations aren’t exactly blazing fast, and are single-threaded (though they will parallelize over shards)."
This sounds extremely pointless with it being single-threaded, but whatever.
Its single-threaded nature for map/reduce is its biggest liability in my opintion. It will apparently be fixed in the next version (https://jira.mongodb.org/browse/SERVER-4258)
Right now, large map/reduce operations just bring the thing to its knees. So we just don't map/reduce with it. The query features are really quite nice, and we do the rest in hypertable.
As Jay Kreps (works at LinkedIn, created Voldemort DB) would wonder, they went from a service managed by Amazon to what?
http://blog.empathybox.com/post/19574936361/getting-real-abo...