Just-In-Time Scalability: Agile Methods to Support Massive Growth

In the course of six months IMVU’s user base quadrupled in size. At the start of this period we were bottlenecked on a single central database. During these six months we evolved IMVU’s architecture to use caching with memcached, replication, horizontal and vertical partitioning to support this growth.

This session will focus on the techniques used to incrementally add scalability without having to make large changes to the application layer or disrupt ongoing feature development by the rest of the team. We’ll look specifically at implementing horizontal partitioning in a way that makes writing scalable application code easy for non-DB experts.

• not all or nothing
• migrate one system or one table at a time
• measure what to scale (based both on slow queries AND frequent queries)
• how to write scripts to measure
• mix of techniques, and when to use which technique (vertical, horizontal, cache, read replication)
• implementing cache coherency (cacheclass), memory based and no disk based
• have a plan, but invest in implementation just ahead of growth
• avoid doing work that is uneccessary (e.g., Ries was absolutely convinced we needed to partition tables tracking virtual currency and virtual goods, yet we still haven’t partitioned them)
• handling large volume usage and reporting data
• “desharding:” pulling views of horizontally partitioned back onto a database for running reports