So now we have wide spectrum of node types and counts for clusters measured on practice. https://instances.vantage.sh/?filter=i3.2xlarge%7Ci3.xlarge%7Ci4i.xlarge%7Cr5ad.2xlarge%7Cr6gd.2xlarge%7Cr6gd.xlarge%7Cx2gd.2xlarge%7Cx2gd.large%7Cx2gd.medium%7Cx2gd.xlarge&cost_duration=daily&compare_on=true As you can see it all have local SSDs.

Real question is: how (and if) can we estimate/derive required cluster capacity from current number of requests and compute, network and disk capacity used?

You can certainly pack multiple stores in a single cluster, and that will almost certainly increase your efficiency, especially if you have a lot of small and medium-sized stores. One of our multi-tenant clusters has ~500 stores in it.

For the request throughput capacity, it's a bit less straightforward and may require some experimentation with the SKUs available to you...

For both of these dimensions, storage and throughput, Venice offers per-store quotas, so that multi-tenancy can be more predictable. i.e. if one store tries to take more than their quota they can be capped, so the other stores are not impacted. There are various strategies you can consider there, with the two main ones being: 1. The sum of the quota you allocate is equal to or less than the capacity of the cluster, or 2. The sum is higher, in which case you are over-provisioning The first one is safer, while the second one is more efficient. Which one is safe enough depends on whether your use cases tend to spike in correlated fashion or independently... if independently, then the over-provisioning approach might be safe enough.