High-Level Design Technical Interview Questions & Answers

Question 1

How do you clarify and scope ambiguous requirements when asked to design a system?

Accepted Answer

When an interviewer asks an intentionally open-ended prompt like 'Design Twitter' or 'Design WhatsApp', they are evaluating whether you blindly jump into architecture or systematically deconstruct ambiguity. The senior approach is to immediately pause, identify the core user journey, define explicit boundary constraints, and establish quantifiable non-functional metrics (such as throughput, latency, and durability) before proposing any components.   - Clarify Features: "Twitter has hundreds of f...

Question 2

What is the difference between Functional and Non-Functional Requirements in System Design?

Accepted Answer

In system design, Functional Requirements define the specific behavioral capabilities and features the software must execute for end users (what the system does), whereas Non-Functional Requirements define the operational qualities, performance characteristics, and constraints under which those features must operate (how well the system performs).   | Dimension | Functional Requirements | Non-Functional Requirements | | :  | :  | :  | | Focus | User actions, inputs, outputs, API contracts | Late...

Question 3

What are SLAs, SLOs, and SLIs, and how do they define system reliability?

Accepted Answer

SLIs, SLOs, and SLAs form the quantitative foundation of Site Reliability Engineering (SRE) used to measure, target, and legally guarantee service performance. An SLI is the empirical metric measured in real time, an SLO is the internal target the engineering team strives to maintain, and an SLA is the contractual business agreement with customers that carries financial penalties if breached.   | Concept | Definition | Example Metric | Audience & Penalty | | :  | :  | :  | :  | | SLI (Indicator)...

Question 4

What is the difference between Availability and Reliability in distributed systems?

Accepted Answer

While often used interchangeably, Availability measures the proportion of time a system is accessible and responds to incoming requests, whereas Reliability measures the probability that a system performs its required function correctly without failure over a specified time interval. A system can be technically available (responding with HTTP 200) while being completely unreliable (returning corrupted data or dropping background jobs).   1. Available but Unreliable:    - An e-commerce search ser...

Question 5

What are the 'Nines of Availability' and what does 99.99% uptime mean in practice?

Accepted Answer

The 'Nines of Availability' is an industry shorthand expressing the percentage of time a service remains operational and accessible over a given period (usually calculated annually). Each additional '9' represents a tenfold reduction in permitted downtime, exponentially increasing architectural complexity, infrastructure redundancy, and operational cost.   | Nines | Availability % | Permitted Downtime / Year | Permitted Downtime / Month | Permitted Downtime / Day | Engineering Architecture | | :...

Question 6

What is the difference between Throughput, Latency, and Response Time?

Accepted Answer

Throughput, Latency, and Response Time represent three distinct performance characteristics of any distributed system. Throughput measures the volume of work a system processes per unit of time (e.g. QPS or bytes/sec), Latency measures the time spent in transit or queueing waiting to be processed, and Response Time measures the complete end-to-end elapsed duration experienced by the client from initiating a request to receiving the final byte.   | Metric | Measurement Unit | Definition | Highway...

Question 7

What are the key latency numbers every system design engineer should know?

Accepted Answer

Originally compiled by Google Senior Fellow Jeff Dean and Peter Norvig, the 'Latency Numbers Every Programmer Should Know' provides the fundamental physical hierarchy of computer operations. In system design, knowing these orders of magnitude allows you to immediately evaluate architectural trade-offs: why caching in RAM is ~100x faster than reading an SSD, and why network round trips dominate all database performance.   1. Memory is 100x Faster Than SSD: A Redis lookup takes ~100 microseconds; ...

Question 8

What is the difference between MTTR, MTTF, and MTBF in incident management?

Accepted Answer

MTTR (Mean Time to Repair/Recover), MTTF (Mean Time to Failure), and MTBF (Mean Time Between Failures) are standard Site Reliability Engineering metrics that quantify how frequently a system breaks and how quickly engineering systems recover. High availability is achieved not just by preventing failures (increasing MTTF), but primarily by automating recovery so failures heal in seconds (minimizing MTTR).   | Metric | Full Form | Formula | Focus | How to Optimize | | :  | :  | :  | :  | :  | | MT...

Question 9

How do you estimate Queries Per Second (QPS) and peak traffic multiplier for a web service?

Accepted Answer

Estimating Queries Per Second (QPS) is the first quantitative step in any System Design interview. It translates high-level business metrics (such as Daily Active Users - DAU) into raw computational engineering load, establishing whether a service can run on a single PostgreSQL instance (handling ~2,000 QPS) or requires a sharded distributed cluster with Redis caches (handling 100,000+ QPS).   - Given: 200 Million Daily Active Users (DAU). - User Activity: Each user views their feed 10 times per...

Question 10

How do you calculate storage capacity requirements for a system over 5 years?

Accepted Answer

Estimating 5-year storage capacity determines whether a system can live on standard block volumes, requires a distributed object store (like AWS S3), or demands tiered cold-storage archiving policies. The calculation requires splitting data into structured relational metadata (lightweight, highly indexed) and unstructured binary media (heavyweight, stored in blob storage), followed by applying multi-year replication and growth factors.   - Scale: 50 Million new photos uploaded per day. - Photo S...

Question 11

How do you estimate network bandwidth requirements for ingress and egress traffic?

Accepted Answer

Network bandwidth estimation establishes how many bits per second transit across your system's ingress (incoming data into servers) and egress (outgoing data sent to clients). Because cloud providers bill heavily for egress bandwidth and physical network interface cards (NICs) have finite caps (e.g. 10 Gbps, 25 Gbps, 100 Gbps), accurate bandwidth modeling dictates whether you need CDNs to offload egress or multi-region routing.   - Scale: 100 Million daily video views. - Video Payload: Average v...

Question 12

How do you calculate memory and cache sizing using the 80/20 Pareto principle?

Accepted Answer

The 80/20 Pareto Principle in system design observes that 80% of read traffic is generated by 20% of the data (the 'hot' data). Because RAM is roughly 100x more expensive per gigabyte than SSD storage, storing 100% of your dataset in a Redis cache is financially irresponsible. By sizing your in-memory cache to hold only that critical 20% of hot daily working data, you can satisfy 80% to 90% of requests directly from memory with sub-millisecond latencies.   - Scale: 200 Million Daily Active Users...

Question 13

How do you perform back-of-the-envelope calculations for read-heavy vs write-heavy workloads?

Accepted Answer

A system's Read-to-Write ratio is the single most decisive factor dictating database selection, caching strategy, and indexing architecture. Systems range from ultra-read-heavy (e.g. Twitter or Wikipedia at 100:1 read ratio) to ultra-write-heavy (e.g. IoT telemetry logging or ride-share GPS tracking at 1:10 read-to-write ratio). Calculating this ratio upfront prevents catastrophic architectural missteps like using B+ Tree SQL databases for write-dominated streaming.   | Dimension | Read-Heavy Ar...

Question 14

What are single points of failure (SPOF) and how do you eliminate them?

Accepted Answer

A Single Point of Failure (SPOF) is any individual component, service, network link, or vendor dependency in an architecture whose failure triggers the total disruption of the entire system. Eliminating SPOFs requires designing N+1 redundancy, automated health detection, and seamless failover mechanisms across every layer of the technology stack (DNS, Load Balancers, Application Servers, Caching, and Storage).   1. Redundancy (N+1): Never deploy an essential service as a single instance. Always ...

Question 15

What is the difference between stateful and stateless service architectures?

Accepted Answer

The distinction between stateful and stateless architectures lies in where client contextual data (such as login sessions, shopping cart contents, or ongoing conversation states) is stored. In a stateful architecture, the server retains client state in its own local memory across consecutive requests; in a stateless architecture, servers retain zero local client state, and every incoming request carries all necessary credentials or queries an external shared storage tier.   | Feature | Stateless...