Database Replication Strategies

Question 1

A client writes a new value to the leader, then immediately reads from a follower and gets the old value. What term best describes this situation?

Accepted Answer

Replication lag — Replication lag is the delay between a write being accepted by the leader and that write appearing on a follower. During this window a follower read can return a stale (old) value, even though the write succeeded. This is an expected trade-off of asynchronous replication, not an error.

Question 2

Why does single-leader replication simplify the write path compared to multi-leader?

Accepted Answer

Only one node accepts writes, so there is never a concurrent write conflict to resolve — With a single leader, all writes are serialised through one node. There is no possibility of two nodes simultaneously accepting conflicting writes to the same key, which eliminates the need for a conflict-resolution strategy. The downside is that write throughput and write availability are capped at whatever one machine can handle.

Question 3

What specific problem does multi-leader replication solve that single-leader cannot?

Accepted Answer

It lets each region accept writes locally, reducing write latency for geographically distributed users — In single-leader replication, a user in Europe writing to a US-hosted leader faces round-trip latency on every write. Multi-leader puts a leader in each region so writes can be accepted locally, dramatically lowering write latency. The cost is that two leaders can receive conflicting writes simultaneously, which must then be reconciled.

Question 4

What new problem does multi-leader replication introduce that single-leader does not have?

Accepted Answer

Write conflicts when two leaders accept different values for the same key at nearly the same time — Because each leader accepts writes independently, it is possible for Leader A to set X=5 and Leader B to set X=9 within the same instant. When they replicate to each other, both are "correct" from their own perspective but they disagree. The system needs a policy — like last-write-wins or application-level merge — to pick or combine the values.

Question 5

A multi-leader system uses "last-write-wins by wall-clock timestamp" to resolve conflicts. What is the key downside of this strategy?

Accepted Answer

One of the conflicting writes is silently discarded, with no indication to the client that their write was lost — Last-write-wins picks the write with the later timestamp and discards the other. The client whose write "lost" receives no error — from their perspective the write succeeded. This can lead to silent data loss, which is often unacceptable in financial, healthcare, or inventory systems where every write must be preserved.

Question 6

In a leaderless (Dynamo-style) system with N=5 nodes, what does a "quorum" of W=3 for writes mean?

Accepted Answer

At least 3 of the 5 nodes must acknowledge the write for it to be considered successful — A quorum for writes (W) is the minimum number of nodes that must acknowledge the write before the client gets a success response. With W=3 and N=5, the write can tolerate up to 2 slow or unreachable nodes and still succeed. The remaining nodes will catch up asynchronously when they come back online.

Question 7

Why does the condition R + W > N matter in a leaderless quorum system?

Accepted Answer

It guarantees that the set of nodes contacted for a read always overlaps with the set that confirmed the last write — If R + W > N, the read quorum and the write quorum must share at least one node. That shared node holds the most recently written value, so the read is guaranteed to see it. For example, N=3, W=2, R=2: R+W=4>3, so at least one node appears in both groups. If R+W were <= N, a read could contact only nodes that missed the latest write.

Question 8

A leaderless system has N=3 nodes. Node 3 is temporarily unreachable. A client writes X=7 with W=2. What happens?

Accepted Answer

The write succeeds once 2 nodes ack; Node 3 will sync the value when it recovers — With W=2 and N=3, only a quorum of 2 acknowledgements is needed. Nodes 1 and 2 both ack, so the write is committed and the client receives success. Node 3 is out of sync but will receive the missing write through an anti-entropy background process or during a subsequent read repair when it comes back online.

Question 9

Which replication strategy offers the highest write availability (least likely to reject writes due to node failures)?

Accepted Answer

Leaderless/Quorum with a small W (e.g. W=1) — because only one node needs to be up — Leaderless/Quorum with a very small W (even W=1) maximises write availability — the write succeeds as long as at least one node is reachable. Multi-leader is also high-availability for writes but requires at least one designated leader per region to be up. Single-leader is the most constrained: if the leader is down, writes must wait for a failover to complete.

Question 10

A social media app stores user profile updates. Users notice they can edit their profile, refresh the page, and occasionally see the old version. Which strategy and mechanism most directly explains this?

Accepted Answer

Leader-follower replication with asynchronous propagation — the read hit a follower that had not yet received the update — This is a classic replication-lag symptom. The write went to the leader successfully, but the follower that served the subsequent page request had not yet received the replicated update. The fix is "read-your-writes" consistency: after a write, route the next read for that user to the leader (or to a follower known to be caught up) for a brief window.

Which replication strategy?

Test your knowledge

A client writes a new value to the leader, then immediately reads from a follower and gets the old value. What term best describes this situation?

The short version

Interview Questions

References & Further Reading

Related topics

Love this visualization?