Sharding Techniques at Mixpanel Engineering

I'm surprised no one has considered using CIDR notation for sharding. If you SHA-256 the key you've got 256-bits of address space. You could store the whole table very compactly and there are plenty of algorithms out there for making the lookup VERY fast.

You could use DNS to transmit the server info eg. do a lookup on a4.c7.3a.d4.ec.1f.5d.f4.c6.f6.5f.77.f3.71.f5.3d.fa.e8.ea.15.cluster.foo.com

In your DNS server just have a record for *.15.cluster.foo.com this effectively gives you a /8 on a 256-bit routing space. To find all the servers you need to query to query the entire keyspace just ask for an AXFR request on cluster.foo.com. This system also makes mirroring and partitioning extremely simple. Say you have disparate servers that can handle disparate amounts of data? Just turn your smaller servers into /9s

Note if you're considering implementing this there are some patent applications around it. Look in my profile if you want to code around the patent application.

This post is a great 101, but it doesn't provide much info on what kind of setup you run at Mixpanel. I see you ditched Cassandra, though, and you mention Riak. Are you using Riak, rolling your own sharding layer, or what?

Step 0 - have enough traction that you need to worry about sharding.