The mapping between the NP, operator, DC, node is the following:
- node has a node_operator field
- node_operator has a dc_id field
- node_operator also has a node_provider field
In result,
- A NP can have [0…many] node operators
- A node operator can have [0…1] DCs
- A DC can have [0…many] node operators
- A node operator can have [0…many] nodes
So yes, the node operator gives us appropriate granularity because the same NP would often (albeit not necessarily always) buy a set of machines with the same specs and from the same supplier, and put these machines into the same rack, in a DC. These machines would also share the power supply and the ISP, so they would be in the same failure domain.
In another DC, we could expect that the NP would use another supplier and another ISP, so it would be a different failure domain.
What is likely wrong in the whole thing is the terminology we use. But someone thought the above naming is “good enough” and now it’s pretty hard to change. Naming is hard.
That said, we can surely start with the NP-based extrapolation (rather than node operator based), and check if that NP-based extrapolation approach would catch the current performance offenders. And if that (NP-based) doesn’t work - switch over to the node operator based extrapolation.