Questions: Replication Strategies and Trade-offs

Question 1

A leaderless replication system has N = 5 nodes. An engineer sets a write quorum of W = 2 and a read quorum of R = 2. What is the fundamental problem with this configuration?

• A. Write quorums must always be a strict majority; W = 2 out of 5 is too small and writes will be rejected
• B. W + R = 4 < N = 5, so it is possible for a read to contact only nodes that missed a recent write, returning stale data without detecting it
• C. This configuration is optimal — W + R < N minimizes latency by reducing coordination overhead
• D. N = 5 requires odd quorum values; even values cause split-brain scenarios during network partitions

Answer: (B)

The quorum overlap guarantee requires W + R > N. With W = 2 and R = 2, W + R = 4 < 5 = N. It is then possible that a write is acknowledged by nodes {1, 2} while a subsequent read contacts nodes {3, 4} — zero overlap, so the read returns stale data without detecting it. Setting W + R > N ensures by the pigeonhole principle that at least one node participated in both the write and the read set, guaranteeing the read sees the latest write. The engineer should increase W to 3 or R to 3 to satisfy W + R > 5.

Question 2

A company uses multi-leader replication so users in the US and Europe can write to their nearest datacenter without cross-continental round trips. A US user and a European user simultaneously update the same product's inventory count. What problem must this system handle that primary-backup replication avoids?

• A. A network partition — multi-leader systems are uniquely vulnerable to split-brain failures not present in primary-backup
• B. A write conflict — both leaders accepted a valid concurrent write to the same record and the system must decide which value wins
• C. A Byzantine failure — one datacenter may send corrupted acknowledgments to the other
• D. A deadlock — both leaders will block indefinitely waiting for the other to release a record lock

Answer: (B)

Primary-backup avoids write conflicts by routing all writes through one node. Multi-leader replication trades this simplicity for lower write latency across geographies, but introduces concurrent conflicting writes. The system must apply a conflict resolution strategy: last-writer-wins (simple but lossy — one update is silently discarded), application-level merging (correct but complex), or CRDTs for specific data structures. There is no universally correct answer — it depends on the application's semantics. This is the fundamental cost of allowing multiple write paths: conflict resolution cannot be avoided, only deferred or automated.

Question 3

In a leaderless replication system, setting W + R > N guarantees that at least one node participating in any successful read has seen the most recent successful write.

• T. True
• F. False

Answer: True

By the pigeonhole principle: if a write was acknowledged by W nodes out of N, and a subsequent read contacts R nodes out of N, then because W + R > N, the write-set and read-set must share at least one node. That overlapping node has the most recent write value and can return it to the client. This is the mathematical foundation of quorum-based consistency. Note the guarantee holds for sequential operations; concurrent writes and network partitions introduce additional complexity that quorum overlap alone cannot fully address.

Question 4

Primary-backup replication offers higher availability than leaderless replication because a single authoritative primary eliminates conflicting writes and ensures correct behavior during node failures.

• T. True
• F. False

Answer: False

Primary-backup actually has *lower* availability in the face of failures. The primary is a single point of failure: when it crashes, the system must pause for failover — elect a new primary — before accepting writes again. Leaderless systems can continue serving reads and writes as long as a quorum of nodes remains reachable, tolerating multiple simultaneous failures without a coordination pause. The tradeoff runs in the opposite direction: primary-backup achieves simpler consistency at the cost of availability during failures, while leaderless achieves higher availability at the cost of more complex consistency management.

Question 5

Why does stronger consistency in a replicated system typically require giving up some availability or accepting higher write latency? Explain the underlying coordination cost.

Model answer: Consistency requires that all replicas agree on the order and content of writes before the system acknowledges success to the client. This agreement requires coordination — nodes must exchange messages confirming a write is durably recorded everywhere before returning. In a synchronous setup, a write cannot complete until the slowest required replica acknowledges it. If any required replica is slow, partitioned, or failed, the write stalls. To avoid stalling, you can either reduce the required quorum (weaker consistency) or accept that writes take longer (higher latency). During a network partition, the CAP theorem formalizes this precisely: you can stay consistent by rejecting writes until the partition heals, or stay available by accepting writes that may diverge, but not both simultaneously.

The core tension is fundamental: consistency is a property about the global state of multiple nodes, but nodes can only communicate through messages that take time and can be lost. Stronger consistency = more coordination = more messages = more opportunities for failure. This is not a design flaw but an inherent property of distributed systems, formalized by the CAP theorem and FLP impossibility result. The right replication strategy is therefore determined entirely by which tradeoffs your application can tolerate.