If you have 100 instances you really want them to share the cache
I think that assumes decoupled compute and storage. If instead I couple compute and storage, I can shard the input, and then I won't share the cache across the instances. I don't think there is one approach that wins every time.
As for egress fees, that is an orthogonal concern.
Comments
It's about scalability. If you have 100 instances you really want them to share the cache so you increase hitrate and keep egress costs low.
I think that assumes decoupled compute and storage. If instead I couple compute and storage, I can shard the input, and then I won't share the cache across the instances. I don't think there is one approach that wins every time.
As for egress fees, that is an orthogonal concern.