Bradley Holt - Scaling CouchDB

Its important to note the distinctions between performance, vertical scaling, and horizontal scaling. Performance typically refers to properties of a system such as response time or throughput. Vertical scaling (or scaling up) means adding computing capacity to a single node in a system. This could be through added memory, a faster CPU, or larger hard drives. While memory gets cheaper, CPUs get faster, and hard drives get larger every year, at any given moment theres an upward limit to vertical scaling. Also, the hardware for high capacity nodes is usually more expensive per unit of computing capacity (as defined by your application) than commodity hardware.

When someone uses the word scalability, they are often referring to horizontal scalability. Horizontal scaling (or scaling out) is when a systems computing capacity is scaled by adding new nodes. In theory, each new node adds the entire amount of that nodes computing capacity to the system. In practice, true horizontal scalability is rarely, if ever, achieved. The network overhead of nodes communicating with each other can detract from the overall computing capacity of the system. Also, few systems are configured such that there is no redundant computing work between nodes.

Capacity planning is often lumped together with scaling. With capacity planning, the focus is on creating a system that can meet an expected amount of demand. With scalability, the focus is on creating a system that is capable of accommodating growth. As mentioned before, both vertical and horizontal scaling each have their own limitations. Given these limitations, a combination of scalability and capacity planning is often warranted. Put another way, define a maximum capacity to which you need to scale and test so that your system can scale to meet that capacitythere is no such thing as infinite scalability, and attempting to create such a system would be prohibitively expensive. In , we will take a look at testing the capacity of your system through distributed load testing.