An interactive study companion for Martin Kleppmann's classic on reliable, scalable, and maintainable systems.
Every system design picks a point on the latency–throughput–consistency triangle. Name your choice explicitly.
A reservation system needs linearizable inventory decrements. A YouTube view counter can be approximate. A voting system needs auditability stronger than either. The 'right' consistency level is domain-specific.
Write down what must never happen (double booking, duplicate vote) before choosing databases. The invariant picks the consistency model.
Instagram tolerates slightly stale like counts (approximate aggregation) but not duplicate posts on retry (idempotent write IDs). Google Spanner pays latency for global consistency where ads billing requires it.
Message ordering within a chat is strict; global search across chats is weaker or absent. The product promise determines per-feature consistency.
// Facebook / Instagram: route reads to primary after posting
async function createPost(userId: string, text: string) {
const post = await db.primary.posts.create({ data: { userId, text } });
// Sticky session: next reads hit primary or caught-up replica
session.set("readFromPrimaryUntil", Date.now() + 5_000);
return post;
}
async function getFeed(userId: string) {
const client = session.get("readFromPrimaryUntil") > Date.now()
? db.primary
: db.replica;
return client.feed.findMany({ where: { userId } });
}