The article discusses a shift in approach for designing distributed systems, moving away from seeking a single, universal consistency model towards a "State-Oriented Consistency." The authors realized their initial mistake was asking "Which consistency model should the cluster use?" instead of "Which consistency guarantee does this specific piece of state actually need?"

This realization stemmed from an incident where pods were killed by the OOM handler due to excessive memory usage. An investigation revealed that on pod boot, session state for the entire fleet was loaded into memory, regardless of which clients would actually reconnect to that specific pod. This was a result of a design that assumed any node should be ready to serve any client, leading to "Uniform Consistency" – applying one strategy everywhere.

The proposed framework involves identifying state, asking what happens if two nodes disagree, finding the minimum required guarantee, and choosing the weakest mechanism to provide it. This process is repeated for each piece of state. The authors emphasize that consistency is a property of individual pieces of state, not the system as a whole. This State-Oriented Consistency approach simplifies design by allowing mechanisms to be chosen based on the specific needs of each state, rather than imposing a blanket rule.