Sharding / partitioning

Replication gave you more read capacity and a way to survive failure, but it has a ceiling. Every replica still holds the entire dataset, and every write still goes through the single leader. So if you have more data than one machine can hold, or more writes than one leader can absorb, replication alone will not save you.

Sharding is the next tool. Instead of copying the whole database, you split it into pieces and put each piece on a different machine. User A's data lives on one shard, user Z's on another. Now the data and the write load are divided across many machines, each responsible for its slice.

That power comes with real complexity: you have to decide how to split the data, watch out for uneven "hotspots", and accept that questions spanning multiple shards get harder. This unit is about making that tradeoff knowingly.