There is a persistent challenge that I keep returning to: how to make disparate nodes function as a single mechanism when the environment around them is as far from ideal as possible.
I am not satisfied with the approach where classic "master-slave" replication is passed off as distributed architecture. It works perfectly in sterile lab conditions, but the moment the network introduces latency and the nodes operate at different speeds, the system begins to crack. We are forced to choose: either we lose data, or we stop writes, or we build complex workarounds on top of the database. This isn't distribution; it is an attempt to trick physics.
I want to build a system that doesn't panic when packets are lost. An architecture where distribution is not an add-on for fault tolerance, but a fundamental property designed for instability from the ground up.…