[dsl]

From my experiences, AWS was pretty much the worst place to run Riak. The node to node network speed was really what killed it.

We replaced a decent number of m1.large instances with three pretty low end boxes with SSDs on a dedicated gigabit switch.

[jeffbarr]

Did you try to run them on Cluster Compute instances within a Placement Group?

[tdfx]

Did you ever happen to do a writeup about your experiences? As someone who is considering a move to Riak, I'd appreciate your insights.

[dsl]

It has been on my todo list, but nothing yet. A few operational notes I can share:

Riak is incredibly awesome for scaling. The only flaw i've found is there is a tipping point where load on the cluster is too high to be able to add a new node and have partitions hand off to it in a reasonable amount of time. You might also run into the same issue if you lose more than one node at time and try to bring up new instances to replace them (every add or remove event results in a rebalance).

JavaScript MapReduce is a joke. I could never find the right combination of incantations to make the VMs perform in a halfway decent manner. We didn't bother trying to implement in Erlang because we didn't have the expertise in house, but supposedly it is a bit better.

If you need multi-site replication, you'll need to build it yourself, or find some other way to work around it. The price tag for paid Riak is well beyond anything I'd consider reasonable. Think more expensive than the hardware you are running it on.

SSDs and 10g ethernet are a must. SoftLayer is probably a better route if you want to be in "the cloud," but definitely consider your own hardware that you can poke with a stick if something seems odd.

[ihsw]

hi1.4xlarge high-IO instances might be of interest to you, however they are vastly larger and more expensive than m1.large instances.