DDIA Learn — Designing Data-Intensive Applications

When data is replicated, reads may not reflect the latest write. Consistency models formalize what clients can expect. The spectrum runs from eventual (replicas converge eventually) to linearizable (behaves like a single copy).

In practice

DynamoDB offers eventual (default) or strong (R+W>N) reads per request. MongoDB readConcern: majority waits for replication to a quorum. Redis async replication means reads from replicas may be stale — many teams read only from the primary for consistency-sensitive data. Caching in Redis amplifies consistency questions: what TTL is acceptable for stale catalog data?

Facebook at scale

News feed tolerates seconds of staleness in Redis cache-aside — a slightly old post order is acceptable. Account balance reads require strong consistency from the primary, not a cached replica.

typescript — Cache-aside with tunable staleness

// Facebook feed cache-aside pattern
async function getFeed(userId: string): Promise<Post[]> {
  const cacheKey = `feed:${userId}`;
  const cached = await redis.get(cacheKey);
  if (cached) return JSON.parse(cached);

  const posts = await db.posts.findMany({
    where: { userId },
    orderBy: { createdAt: "desc" },
    take: 50,
  });
  await redis.setex(cacheKey, 300, JSON.stringify(posts));
  return posts;
}

Consistency Guarantees