For this use case, we are focusing on a distributed computing pattern that integrates a real-time processing platform (that is, Storm) with an analytics engine (that is, Druid); we then pair it with an offline batch processing mechanism (that is, Hadoop) to ensure we have accurate historical metrics.

While that remains the focus, the other key goal we are trying to achieve is continuous availability and fault tolerance. More specifically, the system should tolerate the permanent loss of a node or even a data center. To achieve this kind of availability and fault tolerance, we need to focus a bit more on the persistence.

In a live system, we would use a distributed storage mechanism for persistence, ideally a storage mechanism that supported replication across data centers. Thus, even in a disastrous scenario, whereby a data center is entirely lost, the system can recover without losing data. When interacting with the persistent store, the client will demand...