Replication: The Follow-Up Questions Interviewers Ask, With Answers
Interviewers ask follow-up questions after you mention replication because they want to see how deeply you understand the trade-offs. Replication touches consistency, latency, cost, and operational complexity, so a shallow answer can hide serious gaps.
They also use the topic to gauge how you think about failure modes and scaling. A good candidate will turn a generic statement about “replicating data” into a concrete discussion of quorum reads, cross-region latency, and capacity budgeting.
Consistency Model
1. What consistency guarantees does your replication provide?
I would say our primary replica uses a strongly consistent quorum write (W ≥ ⌊N/2⌋+1) and reads are served from the leader, guaranteeing linearizability for all client operations. For secondary replicas we expose an eventual consistency read-only endpoint, which is useful for analytics workloads that can tolerate stale data. This dual model lets us keep latency low for reads while still offering strict guarantees where the business needs them.
Failure Handling
2. How does the system behave when the leader fails?
When the leader crashes, the remaining nodes run a leader election (e.g., Raft). The election typically finishes within a few hundred milliseconds; we set the election timeout to 500 ms to detect a dead leader quickly without triggering needless elections, while the majority-vote rule is what prevents split-brain. Once a new leader is elected, it re-plays its log to bring any lagging followers up to date. During the election window, writes are rejected with a retryable error, which the client SDK retries with exponential back-off.
Latency Impact
3. What is the latency penalty of replicating across regions?
Cross-region replication adds roughly the round-trip time between data centers. For example, replicating from US-East-1 to EU-West-1 on AWS adds about 80 ms of network latency. If we use synchronous replication for writes, the end-to-end write latency grows from ~30 ms (single region) to ~110 ms. To keep user-visible latency low, we often make writes locally synchronous and propagate them asynchronously to the remote region, accepting a brief window of inconsistency.
Capacity Planning
4. How do you size the number of replicas for a 10 TB dataset with 5 k RPS reads?
Assuming each read request touches 4 KB of data, 5 k RPS translates to 20 MB/s. A typical SSD can sustain ~500 MB/s of sequential reads, so a single replica could handle the load comfortably. However, to provide fault tolerance and load-balancing, we would deploy three replicas. This gives us N=3, quorum W=2, and allows us to survive any single node failure while still handling the full read traffic.
Data Skew
5. What if one shard becomes a hotspot?
If a particular shard receives a disproportionate share of traffic, we would enable dynamic sharding. The system monitors per-shard request rates and can split a hot shard into two smaller shards, redistributing keys based on a consistent-hash ring. For instance, a shard handling 2 k RPS could be split into two 1 k RPS shards, halving the load on each replica and keeping latency within our target of <50 ms.
Write Amplification
6. How does replication affect write amplification and storage costs?
With synchronous replication to three nodes, each logical write incurs three physical writes, increasing storage usage by roughly 3×. If our dataset is 10 TB, we need about 30 TB of raw storage. To mitigate cost, we use compression (e.g., LZ4) that typically reduces size by 2×, bringing the effective storage requirement down to ~15 TB. Additionally, we run compaction jobs during off-peak hours to reclaim space from deleted or overwritten records.
Recovery Time
7. How long does it take to recover a failed replica?
Recovery time depends on the amount of log that needs to be replayed. In our setup, a replica falling behind by 3 hours of writes (≈ 43 GB at 1 k RPS, 4 KB per write) can catch up in about 6 minutes when we stream the log over a 1 Gbps network (≈ 125 MB/s). We also use snapshotting every 6 hours, so a fresh replica can be bootstrapped from the latest snapshot and then apply only the recent log entries.
Consistency vs. Availability Trade-off
8. If you had to choose between stronger consistency and higher availability, what would you pick?
The choice is driven by the business SLA. For a financial transaction service, we prioritize strong consistency even if it means a brief outage during a network partition—writes must not be lost or double-spent. Conversely, for a social feed where eventual consistency is acceptable, we would favor higher availability, allowing reads from any replica and tolerating temporary divergence.
Operational Complexity
9. What operational challenges does replication introduce at scale?
At scale, replication adds several moving parts: leader election, log compaction, network partition handling, and data-center synchronization. Monitoring becomes critical—we need metrics for replication lag, election times, and quorum health. Automated failover scripts must be tested regularly. In addition, version upgrades require rolling restarts that respect quorum constraints, otherwise we risk split-brain scenarios. Investing in robust observability and automated remediation reduces the operational burden.
Senior vs. Staff Answers
A senior-level answer will correctly name the consistency model, quantify latency, and discuss basic failure handling. A staff-level answer goes further: it includes capacity calculations, outlines mitigation strategies for hot shards, and demonstrates awareness of operational tooling and trade-offs across multiple failure domains. Staff candidates also tie the design back to business-level SLAs and show how they would evolve the system as traffic grows.
How to practice this
- Review the worked design of a wide-column store like Cassandra in the System Design AI library; click each component for its notes and look for how replication, consistency levels and failures are handled. It replicates without a single leader, so compare it with the leader-based answers above.
- On the canvas, sketch a replication architecture for a service you care about, labeling quorum sizes, latency budgets, and recovery steps.
- Record yourself answering the nine questions out loud, then compare your responses to the model answers and iterate.
Questions people ask
Why do interviewers focus on replication after a brief mention?
Replication touches many core system properties—consistency, latency, fault tolerance, and cost. Interviewers use follow-up questions to see if you understand those trade-offs and can reason about real-world numbers.
What is the simplest consistency model to explain in an interview?
Start with a strong linearizable model using a quorum write (W ≥ ⌊N/2⌋+1) and leader reads. Then mention an eventual-consistency read-only path for low-latency analytics.
How much extra storage does synchronous replication usually require?
With N replicas, storage grows roughly N-fold. For three replicas, a 10 TB dataset needs about 30 TB of raw storage before compression or deduplication.
What metric should I monitor to detect replication problems?
Track replication lag (the time difference between leader and follower commit timestamps) and quorum health (number of nodes in sync). Sudden spikes in lag often indicate network issues or a failing node.