Scaling Data: Replication, Sharding, and Distributed Transactions

More cashiers, then more shops

Vikram's online grocery, SabziBasket, runs on one PostgreSQL server. It has 2 million customers, and every morning rush pushes the database a little closer to the edge. (Vikram, SabziBasket and every number here are made up for teaching.)

Vikram's father ran a real grocery shop, and he scaled it in two steps:

1. He added a second cashier with a copy of the price list. That is replication: the same data, more hands.
2. When that wasn't enough, he opened a second shop selling half the catalogue. That is sharding: different data on different machines.

Each step solved one problem and created another. The copies of the price list drifted apart, and a customer who wanted items from both shops now needed a plan.

SabziBasket goes through the same steps, and each one breaks something:

This guide follows Inside a database and uses the vocabulary in Consistency models.

1. "Reads are drowning the database": replication

Ninety per cent of SabziBasket's traffic is customers browsing and checking orders. Vikram adds read copies, which are the extra cashiers.

flowchart TD
    subgraph SL["Single leader"]
        W1@{ icon: "algomi:svc-cube", label: "Writes", pos: "b", h: 52 }
        L1@{ icon: "algomi:data-cylinder", label: "Leader", pos: "b", h: 52 }
        W1 --> L1
        F1@{ icon: "algomi:data-cylinder", label: "Follower", pos: "b", h: 52 }
        L1 --> F1
        F2@{ icon: "algomi:data-cylinder", label: "Follower", pos: "b", h: 52 }
        L1 --> F2
    end
    subgraph ML["Multi-leader"]
        W2@{ icon: "algomi:svc-spark", label: "Writes in Mumbai", pos: "b", h: 52 }
        LA@{ icon: "algomi:data-cylinder", label: "Leader Mumbai", pos: "b", h: 52 }
        W2 --> LA
        W3@{ icon: "algomi:svc-cube", label: "Writes in Delhi", pos: "b", h: 52 }
        LB@{ icon: "algomi:data-cylinder", label: "Leader Delhi", pos: "b", h: 52 }
        W3 --> LB
        LA <-->|sync, may conflict| LB
    end
    subgraph LL["Leaderless"]
        W4@{ icon: "algomi:svc-cube", label: "Writes and reads use quorums", pos: "b", h: 52 }
        N1@{ icon: "algomi:data-cylinder", label: "Replica", pos: "b", h: 52 }
        W4 --> N1
        N2@{ icon: "algomi:data-cylinder", label: "Replica", pos: "b", h: 52 }
        W4 --> N2
        N3@{ icon: "algomi:data-cylinder", label: "Replica", pos: "b", h: 52 }
        W4 --> N3
    end