Clustering

A server cluster consists of a set of connected servers that work together and can be seen as a single system. They are usually connected to a fast Local Area Network (LAN). The significant difference between a cluster and a group of servers is that the cluster acts as a single system trying to provide high availability, load balancing, and parallel processing.

If we deploy applications, or services, to individually managed servers and treat them as separate units, the utilization of resources is sub-optimum. We cannot know in advance which group of services should be deployed to a server and utilize resources to their maximum. Moreover, resource usage tends to fluctuate. While, in the morning, some service might require a lot of memory, during the afternoon that usage might be lower. Having predefined servers prevents us from having elasticity that would balance that usage in the best possible way. Even if such a high level of dynamism is not required, predefined servers tend to create problems when something goes wrong, resulting in manual actions to redeploy the affected services to a healthy node:

Real clustering is accomplished when we stop thinking in terms of individual servers and start thinking of a cluster; of all servers as one big entity. That can be better explained if we drop to a bit lower level. When we deploy an application, we tend to specify how much memory or CPU it might need. However, we do not decide which memory slots our application will use nor which CPUs it should utilize. For example, we don’t specify that some application should use CPUs 4, 5 and 7. That would be inefficient and potentially dangerous. We only decide that three CPUs are required. The same approach should be taken on a higher level. We should not care where an application or a service will be deployed but what it needs. We should be able to define that the service has certain requirements and tell some tool to deploy it to whichever server in our cluster, as long as it fulfills the needs we have. The best (if not the only) way to accomplish that is to consider the whole cluster as one entity.

We can increase or decrease the capacity of a cluster by adding or removing servers but, no matter what we do, it should still be a single entity. We define a strategy and let our services be deployed somewhere inside the cluster. Those using cloud providers like Amazon Web Services (AWS), Microsoft Azure and Google Compute Engine (GCE) are already accustomed to this approach, even though they might not be aware of it.

Throughout the rest of this chapter, we'll explore ways to create our cluster and explore tools that can help us with that objective. The fact that we'll be simulating the cluster locally does not mean that the same strategies cannot be applied to public or private clouds and data centers. Quite the opposite: