POJO Mojo

I met with someone last week who has a massively scaled out system. The application contains user / document information. The system is designed for linear scalability today, and this is achieved mostly by hand.

With a series of load balancers and a good partitioning scheme, the architecture delivers about 500 partitions each with fewer than 10,000 users stored inside. This system scales linearly, all on top of a series of relational database instances. The nasty truth is that scale out has become too expensive.

The goal: get more out of each partition. How? By offloading the DB altogether. Detaching from it and keeping the transient data transient--in memory where it belongs.

Have you hit the scalability wall even though your scalability is linear? Tell me more.

--Ari

Using Terracotta on Amazon EC2, Gnip is handling the fire hose coming from many popular social networking sites. Gnip offers a single consolidated and filtered view of all those streams of data.

Gnip uses no DB. It uses Terracotta and backs that up to S3. The most interesting part to me is the throughput. They do more than 50K operations per second to their TC server. Each operation contains 1000 updates in it. This means Gnip is handling 50,000,000 (50 million) updates per second in EC2 on Terracotta.

We have a few engineers writing down what they did so be patient and you will be able to read more detail.

Amazing stuff!

--Ari