It's worth noting that that solution only works if you have much fewer writes than reads, but that's probably true for most people. A really simple way of doing this that all memcached clients should be able to handle is to set up two separate memcached clusters, write to both, and randomly read from either of them. That way you don't replicate nodes, you replicate the entire cluster. :-)
A better solution would be to consolidate your items, i.e. make sure that items that often are fetched together with multi-get end up on the same server, however this requires a lot of extra knowledge about your data, and it's not very likely that this knowledge is available.
> It's worth noting that that solution only works if you have much fewer writes than reads, but that's probably true for most people.
Is it true of tweets? How about e-mail? What facebook content works that way? (IIRC the discussion of their new image store, one of the motivations was that most pictures were never viewed.)
A better solution would be to consolidate your items, i.e. make sure that items that often are fetched together with multi-get end up on the same server, however this requires a lot of extra knowledge about your data, and it's not very likely that this knowledge is available.