Hi, first post to the list for me... Have any folks here considered using the crush mapping algorithm with memcached, for replication?
http://www.ssrc.ucsc.edu/Papers/weil-sc06.pdf It originally comes from the ceph storage project, and it seems like a nice step forward over consistent hashing. The crush alg takes input of a key and a weighted server node hierarchy, instead of a flat server list, The node hierarchy lets crush more closely match physical server reality, representing shelves, racks, rows, etc. For example, you can say "Put replicas onto different shelves than the primary". I'm considering implementing parts of it into libmemcached with some other ideas, but love to hear from the y'all for any wisdom/warnings. ps, nice meeting folks at the hackathon. Was hoping to mention this there, but folks looked too productive on the binary-protocol/multi-engine work and then too tired afterwards. -- steve
