Hot-warm architecture is a concept of having two types of data nodes conceptually (where the hot nodes would probably have better specs)

discovery.zen.minimum_master_nodes = 2 stops split-brain
- from an architecture perspective, master nodes do not handle search or indexing operations so less frequent garbage collection and less required resources
Hot nodes perform all indexing within the cluster
- hold the most recent indices, assumed to be more frequently queried
- indexing is high I/O and CPU, so requires SSD
Warm nodes will be designed to handle large amounts of read-only indices
- read-only implies less computing, typically we need much larger disks (maybe HDD or some ELB)
this should all be transparent to users of the elasticsearch cluster

How do we configure indices to route data to the correct node?

node.attr is the interface exposed by ES for doing this
they suggest the arbitrary attribute box_type or something similar
then, under the index templates' settings, we can specify index.routing.allocation.require.box_type

How do old indices get shuffled from hot to warm? How expensive is this operation?