Question 1

How would you design a globally distributed system to minimize latency for users worldwide?

Accepted Answer

Use anycast routing, deploy PoPs (Points of Presence) in each region, GeoDNS to route users to the nearest region, maintain read replicas regionally, and cache aggressively at the edge. Speed of light: New York → Singapore ≈ 170ms RTT minimum. No amount of optimization beats physical proximity. CDN edge serves static content. Regional read replicas serve dynamic reads locally. Anycast routes BGP traffic to the nearest PoP. GeoDNS resolves to the nearest region's IP.

Question 2

What is mTLS (Mutual TLS) and when is it used in microservices architectures?

Accepted Answer

Both client and server authenticate each other using TLS certificates — used in service meshes (Istio, Linkerd) to ensure only authorized services can communicate, providing zero-trust network security between microservices. Standard TLS: server proves identity; client trusts it. mTLS: both prove identity. In a service mesh, each sidecar proxy (Envoy) handles mTLS transparently. Services only communicate if both hold a valid certificate from the same CA. Even if an attacker gets inside your network, they can't impersonate a service.

Question 3

How would you design a pub/sub messaging system to handle 1 million messages per second?

Accepted Answer

Use Apache Kafka with multiple partitions and consumer groups — partition by topic/key for parallelism, configure replication factor for durability, use async producers with batching, and tune linger.ms and batch.size for throughput. Kafka: topics partitioned across brokers. Producers write to partition leaders. Consumer groups read in parallel — one consumer per partition. Key insight: partition count = maximum parallelism. At 1M msg/s: 100 partitions × 10K msg/s each. Tune: batch.size=16KB, linger.ms=5, compression.type=lz4.

Question 4

What is the difference between Layer 4 and Layer 7 load balancing?

Accepted Answer

Layer 4 (TCP/UDP) routes based on IP and port without inspecting content — lower latency, less CPU; Layer 7 (HTTP) routes based on URL paths, headers, or cookies — enables advanced routing (A/B testing, canary, microservice routing) but adds overhead. L4 (AWS NLB, HAProxy TCP mode): terminate TCP at the LB and forward to backends — millions of connections, lowest latency. L7 (AWS ALB, nginx): parse HTTP — route /api → service A, /web → service B, sticky sessions by cookie, SSL termination, request rewriting. Most production setups use L4 for raw throughput + L7 for smart routing.

Question 5

What is a service mesh and what problem does it solve?

Accepted Answer

An infrastructure layer (Istio, Linkerd, Consul Connect) that handles service-to-service communication concerns — mTLS, load balancing, circuit breaking, retries, observability — as a sidecar proxy, removing this logic from application code. Without service mesh: every service implements retries, timeouts, circuit breakers, mTLS, tracing. With Istio: inject Envoy sidecar into every pod — it handles all of this transparently. Application code focuses on business logic. Trade-off: operational complexity, latency overhead (~1ms per hop), steep learning curve.

Computer Networks — Tech Lead Interview Questions

More Tech Lead Topics

Computer Networks at Other Levels