We had to "spread" the load on the key, which often gets on memcache servers. We do this by simply adding a random number 0-N, where N is a multiple of the number of instances you have configured. If you miss, you will go to the source (database, whatever). I say several times, since memcache hashes are against your key, you may get a bucketing collision without knowing.
This can lead to an increase in reading costs from the source (proportional to your N edge), but it allows you to distribute the load by one key in all copies, keeping these boxes happy.
But yes, this is not for redundancy, but for load balancing.
source share