A node may hold more or less than that depending on the data, mappings and workload. These are all rough guidelines around best practices, and I created this blog post as I saw a lot of users ending up in serious trouble due to having large volumes of small shards, which for many use cases is very inefficient.