Design a sharded, multi-region key-value store with linearizable reads and a 99.99% availability SLO across two Atlantic regions. Discuss the failure modes you would expect during a transatlantic link partition, the consensus mechanism you would use for the metadata plane, and the operational tools you would build to detect and recover from a stale-shard scenario.

Okay, before I touch the data plane I want to be explicit about what 'linearizable' means here — every read sees a value at least as fresh as the most recent successful write. That tells me I cannot serve reads from a region currently fenced off from the leader. Step one: which region holds the leader for each shard? I want a metadata plane that's smaller, separately replicated, and uses Raft so leadership is unambiguous. Now the failure modes — partition between Atlantic regions is the headline one. If we're synchronous, every write blocks during a partition. That's where the SLO bites: 99.99% means about 52 minutes of total annual downtime…