Build your own wide-column store (Cassandra / DynamoDB family)

You have run a single-node KV or SQL database. Maybe you have watched the disk fill. Maybe you have watched throughput peg the ceiling. Maybe you have watched the box die and taken your data with it. The single server has three ways to die — and bigger disks, faster CPUs, and more RAM make those three problems quantitatively softer, but never qualitatively softer. To survive any of them, you need more than one machine. That is the brief.

The wide-column store is the family of distributed databases that took that brief seriously: Cassandra, DynamoDB, ScyllaDB, Riak. They share a load-bearing design — Dynamo-style hash partitioning, tunable consistency, anti-entropy — that is so widely deployed it is the default mental model for "database that survives a node death." Once you have built one from first principles, every later distributed database collapses to a point in this design space.

Resist the urge to memorize Cassandra's command-line flags. Build the system from the failure that motivates each step. The first server fails three ways; the cheapest fix shards by key; sharding by hash mod N reshuffles 80% of keys when you add a server; the ring fixes it but creates uneven arcs; vnodes fix the unevenness; copies survive a node death; copies that disagree need a winner; counting replicas (W+R>N) gives you reads that see writes; partitions force a choice; hinted handoff softens the everyday case; read repair and anti-entropy heal the rest; gossip tells the cluster who is alive. Twelve scenes; twelve choices; one design canvas at the end.