System Design — Top Questions at MAANG (2026)

At L4/SDE-II/E4 the bar is: complete reqs → high-level → 1 deep-dive on a subsystem → discuss trade-offs. You have 45–60 min. 2026 shift: ~50% of SD loops now include an ML/GenAI-adjacent design (was ~10% in 2024). You already have real LLM rate-limiter and EHR-summarization experience — leverage it.

1. Classic top 12 (still asked every cycle)

#	Design	Key concepts	Most-asked at
1	URL Shortener (TinyURL / Bitly)	Base62 encoding, KGS (key gen service), cache, sharding by hash, 301 vs 302	Amazon, Apple
2	News Feed (Twitter / Instagram)	Push vs pull vs hybrid, fan-out, ranking, timeline cache, hot users problem	Meta ★★★
3	WhatsApp / Messenger	Long-poll vs WebSocket vs MQTT, end-to-end encryption, online presence, group fan-out, message ordering	Meta, Amazon
4	Uber / Lyft (ride-matching)	Geo-hashing, S2/H3, driver location stream, matching service, surge pricing	Google, Amazon
5	YouTube / Netflix (video)	Transcoding pipeline (you’ve built one!), CDN, ABR streaming (HLS/DASH), recommendation engine	Netflix ★★★, Amazon
6	Google Drive / Dropbox	Chunking, dedup, sync algorithm, conflict resolution, metadata service	Google, Amazon
7	Search engine / autocomplete	Inverted index, trie+heap for autocomplete, prefix freq, sharded index	Google ★★★, Amazon
8	Distributed cache (Redis cluster)	Consistent hashing, replication (sync/async), eviction (LRU/LFU), hot key problem, cache stampede	Amazon, Meta
9	Rate limiter	Fixed window / sliding window / token bucket / leaky bucket, Redis-based vs in-memory, distributed sync	Amazon ★★, Stripe
10	Notification system	Multi-channel (push/SMS/email), priority queue, retry with backoff, dedup, fan-out	Amazon
11	Web crawler	BFS frontier, politeness (robots.txt + per-host rate limit), URL dedup (bloom filter), distributed workers	Google
12	Online code editor (Leetcode/Coderpad)	Sandboxed containers, persistent storage of submissions, real-time collab (OT/CRDT)	Coderpad, Meta

2. 2026 NEW: GenAI / LLM-infra designs (HIGH PRIORITY for you)

These now appear in ~50% of senior loops. You have direct production experience — frame your Jio work in interview language.

#	Design	Frame using your Jio work	Companies asking
13	LLM API serving (think: design ChatGPT API backend)	Your AI Rate Limiter + LLM Orchestration	OpenAI, Anthropic, Google, Meta, AWS Bedrock teams
14	RAG system	Vector DB (Pinecone / pgvector / Weaviate), embedding pipeline, hybrid retrieval, re-ranking, prompt cache	Cohere, Glean, every AI startup
15	Multi-tenant LLM cost optimizer / rate limiter	This IS your project — write it up as gold-standard answer	Amazon, Anthropic, Meta AI infra
16	Real-time clinical AI summarization	Your Clinical AI Summarization Engine — FHIR ingest → batching → LLM → cache → guardrails	Verily, Tempus, Hippocratic AI, Anthropic Health
17	AI agent backend (Claude-Code / Cursor style)	Tool-calling protocol, sandboxed exec, conversation memory, streaming, observability	Anthropic, OpenAI, Cursor
18	Vector search at scale	Approx NN (HNSW, IVF, ScaNN), sharding, hybrid index, freshness	Pinecone, Meta, Google
19	Feature store for ML	Online + offline parity, point-in-time correctness, registries	Tecton, Uber, Netflix

Interviewer ask in 2026: “How would you design serving infra for an LLM API that handles thousands of concurrent requests with variable latencies?” — answer with your AI Rate Limiter as the concrete example you’d build on.

3. Healthcare-domain SD (warm rounds at HealthTech)

Use these at Verily / Oscar / Tempus / Cigna-Evernorth / Hippocratic AI / Aetna / Walmart Health.

Design HIPAA-compliant EHR data sync (CDC + Kafka + Postgres — your CDC microservice)
Design FHIR resource server (RESTful, eventual consistency, paging, $search, OAuth/SMART-on-FHIR)
Design e-Prescribing system (idempotency, audit, NCPDP integration)
Design real-time medical-alerts pipeline (vitals stream → ML scoring → on-call notification)
Design patient-reported outcomes (ePRO) collection at scale — you’ve shipped this

4. The 6-step framework (use this religiously in every SD interview)

1. Clarify       (2–5 min)  Functional + non-functional reqs, scale, read/write ratio, latency target
2. API           (3 min)    REST/gRPC signatures, key entities
3. Data model    (3 min)    Tables/collections, indices, partition keys
4. High-level    (10 min)   Boxes & arrows: client → LB → service → cache → DB + async (Kafka)
5. Deep-dive     (15 min)   ONE component the interviewer picks (or you pick the spiciest)
6. Trade-offs    (5 min)    CAP, sync vs async, SQL vs NoSQL, push vs pull, monolith vs MS

Buffer 5 min for questions. Practice the boxes-and-arrows in 10 min rule — it’s the #1 reason candidates fail.

5. Numbers you MUST have memorized (back-of-envelope)

Op	Latency
L1 cache	0.5 ns
L2	7 ns
Main memory	100 ns
SSD random read	150 µs
Round-trip same DC	0.5 ms
Round-trip cross-region	100–150 ms
HDD seek	10 ms
Read 1 MB from memory	250 µs
Read 1 MB from SSD	1 ms
Read 1 MB from network	10 ms
Read 1 MB from HDD	30 ms

QPS / storage approximations:

1 day = 86 400 sec ≈ 10⁵
1 million writes/day ≈ 12 writes/sec, with peak ×3 = 36 writes/sec
1 user post avg 500 bytes; 1 B users × 1 post/day = 500 GB/day
Single MySQL box: ~5–10 K reads/sec, ~1–2 K writes/sec
Single Redis box: ~100 K ops/sec
1 Kafka partition: ~10 MB/s, ~1 K msg/s sustainable

6. Trade-off cheatsheet (memorize the words)

CAP: pick 2. Cassandra/Dynamo = AP. HBase/MongoDB-default = CP. PACELC is the better mental model.
SQL vs NoSQL: SQL → ACID, joins, complex queries. NoSQL → horizontal scale, schema flex.
Push vs Pull (feed): push = fast read, write amp. Pull = light write, slow read. Hybrid for celebrities.
Sync vs async replication: sync = strong consistency + write latency; async = eventual + risk of data loss.
Caching strategies: cache-aside (Redis next to DB), write-through, write-back, write-around.
Idempotency: client-supplied request ID + dedup table — talk about this in any write API.
Backpressure: explicit token bucket > implicit thread pool exhaustion.

7. Practice schedule

Week	Designs
10	TinyURL, Twitter, WhatsApp
11	Uber, YouTube/Netflix, Instagram
12	Google Drive, Distributed cache, Rate limiter
13	LLM serving, RAG, Cursor-backend (GenAI focus)
14	FHIR server, EHR sync, ePRO pipeline (healthcare)

For each: write a single markdown file in this folder with:

Functional + non-functional reqs
API
Data model
Diagram (ASCII or describe)
1 deep-dive
3 trade-offs you’d flag