Before we look at how to keep a cluster running, let's explore some aspects of setting it up in the first place.
When considering a new Hadoop cluster, one of the first questions is how much capacity to start with. We know that we can add additional nodes as our needs grow, but we also want to start off in a way that eases that growth.
There really is no clear-cut answer here, as it will depend largely on the size of the data sets you will be processing and the complexity of the jobs to be executed. The only near-absolute is to say that if you want a replication factor of n, you should have at least that many nodes. Remember though that nodes will fail, and if you have the same number of nodes as the default replication factor then any single failure will push blocks into an under-replicated state. In most clusters with tens or hundreds of nodes, this is not a concern; but for very small clusters with a replication factor of 3, the safest approach would...