I didn’t say you should only have 1 shard per data node. It’s just about slicing your indices into smaller shards to make reallocation and recovery easier and faster. Here, I have some clusters with more than 10k indices and 10B documents (plus the replica) over 20 data nodes and that works smoothly.
Regarding running queries on the master nodes, I’d say “it depends”. If you’re running bulk searches and indexing, then… don’t. I’ve lost enough clusters because people were running bulks on the masters and the masters didn’t have enough resources to both manage the 10k indices and compute the ququeries
Thank you for your comments, it proves some things were not clear, I’ll update the guide.