Hello Dmitry, This is a follow-up on our today's discussion.
I do think you can make a client-only clustering solution with automatic rebalance when a node is added or deleted. You need to use the standard technique with a consistent hash ring and virtual nodes. You also need to maintain data redundancy. The technique is described, for example, here: www.christof-strauch.de/nosqldbs.pdf (in this paper you can also find a link to the original paper). In a nutshell, you create a consistent hash ring of, say, 10k virtual servers. Each virtual server is mapped to a physical server: - if, say, there are 10 physical servers, and redundancy factory is 2, server 1 contains virtual servers 1-1000, plus copies of virtual servers 9000-1000, server 2 contains data virtual servers 2000-3000 plus copies of virtual servers 8000-9000. Actually, which virtual servers are replicated on which physical server can be also calculated using a hash function (essentially, another hash ring). Then your client can have the following operations: - add physical server: this operation only adds the server to the mapping, i.e. assigns a bunch of virtual servers, to it, and that's all. - remove the physical server. This operation removes the server from the consistent hash ring, and redistributes the keys assigned to its virtual servers among adjacent virtual servers on the hash ring. (the keys can be looked up on physical servers contain replicas of the virtual servers which are gone). This redistribution needs to happen to keep the redundancy factor. - insert a key -- this operation inserts the key into the correct virtual, and eventually physical server, and to replicas. - find a key - this tries to find the key in the server according to the current state of the consistent hash ring. If the key is not found, it checks replicas. If the key is not found on replicas, it checks the key on the physical server which is responsible for the next range of virtual servers in the hash ring, until the key is found. Here, of course, we can end up with a lot of checks for missing keys: but this can later on be optimized by keeping information about whether or not the configuration is in a degraded state, and only doing the extra checks on degraded virtual servers. -- _______________________________________________ Mailing list: https://launchpad.net/~tarantool-developers Post to : [email protected] Unsubscribe : https://launchpad.net/~tarantool-developers More help : https://help.launchpad.net/ListHelp
