While a big number of users use large farms of Redis nodes, from the point of view of the project itself currently Redis is a mostly single-instance business. I've big plans about going distributed with the project, to the extent that I'm no longer evaluating any threaded version of Redis: for me from the point of view of Redis a core is like a computer, so that scaling multi core or on a cluster of computers is the same conceptually. Multiple instances is a share-nothing architecture. Everything makes sense AS LONG AS we have a *credible way* to shard :-) This is why Redis Cluster will be the main focus of 2013 for Redis, and finally, now that Redis 2.6 is out and is showing to be pretty stable and mature, it is the right moment to focus on Redis Cluster, Redis Sentinel, and other long awaited improvements in the area of replication (partial resynchronization). However the reality is that Redis Cluster is not yet production ready and requires months of work. Still our users already need to shard data on multiple instances in order to distribute the load, and especially in order to use many computers to get a big amount of RAM ready for data. The sole option so far was client side sharding. Client side sharding has advantages as there are no intermediate layers between clients and nodes, nor routing of request, so it is a very scalable setup (linearly scalable, basically). However to implement it reliably requires some tuning, a way to take clients configuration in sync, and the availability of a solid client with consistent hashing support or some other partitioning algorithm. Apparently there is a big news in the landscape, and has something to do with Twitter, where one of the biggest Redis farms deployed happen to serve timelines to users. So it comes as no surprise that the project I'm talking about in this blog post comes from the Twitter Open Source division. Twemproxy --- Twemproxy is a fast single-threaded proxy supporting the Memcached ASCII protocol and more recently the Redis protocol: https://github.com/twitter/twemproxy It is written entirely in C and is licensed under the Apache 2.0 License. The project works on Linux and AFAIK can't be compiled on OSX because it relies on the epoll API. I did my tests using my Ubuntu 12.04 desktop. But well, I'm still not saying anything useful. What twemproxy does actually? (Note: I'll focus on the Redis part, but the project is also able to do the same things for memcached as well). 1) It works as a proxy between your clients and many Redis instances. 2) It is able to automatically shard data among the configured Redis instances. 3) It supports consistent hashing with different strategies and hashing functions. What's awesome about Twemproxy is that it can be configured both to disable nodes on failure, and retry after some time, or to stick to the specified keys -> servers map. This means that it is suitable both for sharding a Redis data set when Redis is used as a data store (disabling the node ejection), and when Redis is using as a cache, enabling node-ejection for cheap (as in simple, not as in bad quality) high availability. The bottom line here is: if you enable node-ejection your data may end into other nodes when a node fails, so there is no guarantee about consistency. On the other side if you disable node-ejection you need to have a per-instance high availability setup, for example using automatic failover via Redis Sentinel. Installation --- Before diving more inside the project features, I've good news, it is trivial to build on Linux. Well, not as trivial as Redis, but… you just need to follow those simple steps: apt-get install automake apt-get install libtool git clone git://github.com/twitter/twemproxy.git cd twemproxy autoreconf -fvi ./configure --enable-debug=log make src/nutcracker -h It is pretty trivial to configure as well, and there is sufficient documentation in the project github page to have a smooth first experience. For instance I used the following configuration: redis1: listen: 0.0.0.0:9999 redis: true hash: fnv1a_64 distribution: ketama auto_eject_hosts: true timeout: 400 server_retry_timeout: 2000 server_failure_limit: 1 servers: - 127.0.0.1:6379:1 - 127.0.0.1:6380:1 - 127.0.0.1:6381:1 - 127.0.0.1:6382:1 redis2: listen: 0.0.0.0:10000 redis: true hash: fnv1a_64 distribution: ketama auto_eject_hosts: false timeout: 400 servers: - 127.0.0.1:6379:1 - 127.0.0.1:6380:1 - 127.0.0.1:6381:1 - 127.0.0.1:6382:1 Basically the first cluster is configured with node ejection, and the second as a static map among the configured instances. What is great is that you can have multiple setups at the same time possibly involving the same hosts. However for production I find more appropriate to use multiple instances to use multiple cores. Single point of failure? --- Another very interesting thing is that, actually, using this setup does not mean you have a single point of failure, since you can run multiple instances of twemproxy and let your client connect to the first available. Basically what you are doing with twemproxy is to separate the sharding logic from your client. At this point a basic client will do the trick, sharding will be handled by the proxy. It is a straightforward but safe approach to partitioning IMHO. Currently that Redis Cluster is not available, I would say, it is the way to go for most users that want a cluster of Redis instances today. But read about the limitations before to get too excited ;) Limitations --- I think that twemproxy do it right, not supporting multiple keys commands nor transactions. Currently is AFAIK even more strict than Redis Cluster that instead allows MULTI/EXEC blocks if all the commands are about the same key. But IMHO it's the way to go, distribute the subset you can distribute efficiently, and pose this as a design challenge early to the user, instead to invest a big amount of resources into "just works" implementations that try to aggregate data from multiple instances, but that will hardly be fast enough once you start to have serious loads because of too big constant times to move data around. However there is some support for commands with multiple keys. MGET and DEL are handled correctly. Interestingly MGET will split the request among different servers and will return the reply as a single entity. This is pretty cool even if I don't get the right performance numbers with this feature (see later). Anyway the fact that multi-key commands and transactions are not supported it means that twemproxy is not for everybody, exactly like Redis Cluster itself. Especially since apparently EVAL is not supported (I think they should support it! It's trivial, EVAL is designed to work in a proxy like that because key names are explicit). Things that could be improved --- Error reporting is not always stellar. Sending a non supported command closes the connection. Similarly sending just a "GET" from redis-cli does not report any error about bad number of arguments but hangs the connection forever. However other errors from the server are passed to the client correctly: redis metal:10000> get list (error) WRONGTYPE Operation against a key holding the wrong kind of value Another thing that I would love to see is support for automatic failover. There are many alternatives: 1) twemproxy is already able to monitor instance errors, count the number of errors, and eject the node when enough errors are detected. Well it is a shame it is not able to take slave nodes as alternatives, and instead of eject nodes use the alternate nodes just after sending a SLAVE OF NOONE command. This would turn it into an HA solution as well. 2) Or alternatively, I would love if it could be able to work in tandem with Redis Sentinel, checking the Sentinel configuration regularly to upgrade the servers table if a failover happened. 3) Another alternative is to provide a way to hot-configure twemproxy so that on fail overs Sentinel could switch the configuration of the proxy ASAP. There are many alternatives, but basically, some support for HA could be great. Performances --- This Thing Is Fast. Really fast, it is almost as fast as talking directly with Redis. I would say you lose 20% of performances at worst. My only issue with performances is that IMHO MGET could use some improvement when the command is distributed among instances. After all if the proxy has similar latency between it and all the Redis instances (very likely), if the MGETs are sent at the same time, likely the replies will reach the proxy about at the same time. So I expected to see almost the same numbers with an MGET as I see when I run the MGET against a single instance, but I get only 50% of the operations per second. Maybe it's the time to reconstruct the reply, I'm not sure. Conclusions --- It is a great project, and since Redis Cluster is yet not here, I strongly suggest Redis users to give it a try. Personally I'm going to link it in some visible place in the Redis project site. I think the Twitter guys here provided some real value to Redis itself with their project, so… Kudos!