On Sun, Mar 20, 2011 at 8:12 PM, Anurag Priyam <[email protected]> wrote: [...] > I am interested in implementing native sharding support in libdrizzle, > and would like to take it up as a summer of code project. I want to
Here is an overview of what I am thinking. ***** Overview ======== The idea behind sharding is to split data across multiple nodes in an order preserving manner. Queries are therefore distributed across various nodes. Mapping Queries to Shards ------------------------- Use a vBucket[1][2] style partitioning scheme coupled with libhashkit[3] for a pluggable hashing solution. vBucket offers greater flexibility over ketama in partitioning data. It follows a two stage mapping: a. a hashing fucntion maps data to virtual buckets b. virtual buckets statically map to nodes This scheme lets one effectively manage hotspots, and easily add/remove nodes over any cluster size. Libhashkit lets one choose from a number of pre-defined hashing functions, or define one of your own, which can be leveraged to control the spread of data across virtual buckets. Shard groups --------------- For a vBucket we can have an entry in a vBucket map that is an array of indices into a server list. Something like: server_list = ['yeban.in:3306', 'yeban.in:3307', 'yeban.in:3308'] vbucket_map = [ [0, 1, 2], [1, 2, 0], [2, 1, 0] ] In the above configuration, each server is capable of serving a vbucket. vBucket has the concept of states. Typically (as used in libvbucket[4]), a vBucket can have one of the following states: a. Active - This server is servicing all requests for this vbucket. b. Dead - This server is not in any way responsible for this vbucket c. Replica - Handle replication but do not serve. d. Pending - Do not serve. Since, vBucket is the acutal partition entity, and more than one server can handle a vBucket, we can have shard groups. The responsibility of a member within a shard group can be controlled by modifying states. Shard Group Topology -------------------- Drizzle has a pluggable replication system, so that one can write a replication scheme of varying complextity. For eg: a. The default slave plugin's one to one replication[5] a. Joe Daly's Heirearchial Replication[6] b. David Shrewsbury's Master Master Replication[7] Let us slightly modify the default semantics of a vBucket state, so that, a replica can server read requests. It is then possible to setup various shard group topologies, ranging from simple master-slave, to master-master, to multiple master - multiple slave, depending on an applications requirements. A simple master-slave scheme would be something like: a. for a given vBucket, we have one active, and several replicas b. write to the active server, which replicates to the replicas c. read from the replicas Handling Failure ---------------- In the event of the failure of an active node, we change the state of one of the replicas to active. Addition/Removal of New Servers ------------------------------- On addition/removal of a server we need to migrate vBuckets from one server to the other. To add a new server: a. map a vBucket to the new server (add it to a group) b. mark the new one as dead c. replicate to the new server d. mark the new server active, and the old one dead, or replica Implementation ============== TODO Testing ======= TODO Andrew was suggestive of providing access to Jenkins build slaves. About me ======== TODO Mentor ===== Andrew, Brian, Stewart? Links ===== [1] http://techzone.couchbase.com/wiki/display/membase/vBuckets [2] http://dustin.github.com/2010/06/29/memcached-vbuckets.html [3] http://bazaar.launchpad.net/~libmemcached-developers/libmemcached/trunk/files/head:/libhashkit/ [4] https://github.com/membase/libvbucket [5] http://docs.drizzle.org/plugins/slave/index.html [6] http://www.8bitsofbytes.com/?p=28 [7] http://dshrewsbury.blogspot.com/2011/03/multi-master-support-in-drizzle.html ***** -- Anurag Priyam http://about.me/yeban/ _______________________________________________ Mailing list: https://launchpad.net/~drizzle-discuss Post to : [email protected] Unsubscribe : https://launchpad.net/~drizzle-discuss More help : https://help.launchpad.net/ListHelp
