We have to decide the number of shards at the time of creating the index. The number of shards cannot be changed once the index has been created. There is no golden rule that will help you decide how many shards should be created at the time of creating an index. The number of shards actually decides the level of parallelism in the index. Let's understand this by taking an example of how a search query might be executed.

When a search or aggregation query is sent by a client, it is first received by one of the nodes in the cluster. That node acts as a coordinator for that request. The coordinating node sends requests to all the shards on the cluster and waits for the response from all shards. Once the response is received by the coordinating node from all shards, it collates the response and sends it back to the original client.

What this means is, when we have a greater number of shards, each shard has to do relatively less work and parallelism can be increased. 

But can we choose an arbitrarily big number of shards? Let's look at this in the next couple of sub-sections.