System Design
Interview Guide

The core challenge is operational transformation (OT) or CRDTs for conflict-free merges. At L6, go deep on the sync protocol — how does a client reconnect after being offline for 3 days? How do you handle edits that conflict at the character level?

Go deep on the transcoding pipeline — parallelisation, codec selection, adaptive bitrate streaming (DASH/HLS). At L6, discuss CDN strategy, cache warming for viral videos, and how to handle a video that gets 10M views in the first hour.

The key tension is accuracy vs latency. Centralised counters are accurate but slow. Local counters are fast but allow overage. At L6, discuss the token bucket vs sliding window algorithms, Redis cluster strategies, and how to handle clock skew across datacenters.

Go deep on the politeness protocol — how do you avoid overloading small websites while crawling at massive scale? At L6, discuss the URL frontier prioritisation, duplicate detection via SimHash/MinHash, and the incremental indexing problem for frequently updated pages.

At E6, don’t just describe a fan-out architecture. Go deep on the ranking model serving infrastructure — how do you serve personalised ML rankings at 3B users with <100ms p99 latency? Discuss the push vs pull trade-off for different user types (celebrities vs regular users).

Go deep on the message delivery guarantee model — at-most-once vs at-least-once vs exactly-once. At E6, discuss the Signal Protocol for E2E encryption and how key distribution works at scale. How do you handle a user who’s offline for 30 days?

Focus on the Haystack object storage system (Meta’s real solution). At E6, go beyond “use S3” — discuss content-addressable storage, deduplication, CDN strategy for different content types (profile pics vs Reels), and storage cost optimisation through tiering.

A genuine breadth question touching currency handling, regulatory compliance, fraud detection, latency, and reliability. At L7, identify which of these is the hardest problem and go deep on it — don’t cover all five shallowly. Frame decisions using Leadership Principles: “Customer Obsession” for fraud UX, “Operational Excellence” for payment retries.

Amazon will probe what happens when a job is partially executed and a worker crashes. At L6, go deep on exactly-once semantics — idempotent job design, distributed locks via ZooKeeper/DynamoDB, and the at-least-once vs exactly-once trade-off. How do you handle a job that’s been running for 6 hours when the worker dies?

At L6, think about the full loop: offline training → feature store → real-time serving → A/B testing → feedback loop. The hardest part is the feature store — how do you serve freshly-computed features (last purchase 5 minutes ago) at millisecond latency with 100% freshness?

The core challenge is conflict resolution with end-to-end encryption — Apple can’t see your files, so the server can’t resolve conflicts. At ICT6, go deep on vector clocks for conflict detection, and how you do deduplication when you can’t read the content. Also discuss the bandwidth optimisation for mobile clients with limited data plans.

Apple’s on-device ML is a core differentiator. At ICT6, discuss model quantisation for on-device inference, the latency vs accuracy trade-off between on-device and cloud models, and how to use Private Cloud Compute — Apple’s privacy-preserving cloud inference using TEEs (Trusted Execution Environments).

Apple Pay uses device account numbers (DANs) — Apple never sends your actual card number to the merchant. At ICT6, go deep on the tokenisation scheme, the Secure Element on the device, and how the NFC transaction flow works with biometric authentication. Privacy constraint: Apple doesn’t track what you buy.

The core challenge is the presence system at scale — knowing if 300M users are online/offline/in-a-meeting in real time. At L65, go deep on the WebSocket connection management, presence heartbeat optimisation, and how you handle enterprise compliance requirements (message archival, eDiscovery, data residency).

This is about LLM inference at scale with very tight latency constraints (<200ms for inline suggestions). At L65, discuss speculative decoding, KV cache sharing across requests, and the routing strategy for different model sizes (small model for fast suggestions vs large model for complex completions). Also discuss the privacy challenge — code is sensitive IP.

Multi-tenancy is the defining challenge in Azure. At L65, go deep on tenant isolation levels (shared infrastructure vs dedicated), noisy neighbour prevention, resource quota enforcement, and how you handle a tenant that suddenly needs 100x their normal resources (burst capacity). Discuss the billing and metering infrastructure too.