Company by company, LLD and HLD. Sources are 2025–2026 only — older reports were found and discarded, because the format changed materially: Karat is gone at Atlassian, Meta's Product Architecture became AI-enabled in April 2026, and Google is piloting a code-comprehension round.
Both rounds, both real. The hardest machine-coding round of your targets.
Recruiter → CodeSignal OA (70–90 min, 4 problems) → phone screen → 4–6 round onsite → hiring committee. Two to six weeks end to end.
Yes — dedicated, now standard for SDE-2. 45–60 minutes.
The code is compiled and executed. It must run. Judged on data structures, design patterns, language proficiency, single-responsibility decomposition, naming and readability — with extensibility weighted heavily.
Yes — a separate round, even at L4.
City-scale systems with event streams, time windows, hot partitions and correctness trade-offs. They probe real-world large-scale trade-offs rather than implementation detail: clarify ambiguous requirements, decompose into services, pick data models and APIs, reason about failure modes, defend consistency versus availability.
The most exploitable loop you are targeting.
Karat has been dropped. The Coding Design stage is now two rounds — one DSA, one LLD — with system design also appearing in the screen.
Yes, and the format has a signature move.
A mid-round scale-up requirement. They expect a design where the new requirement costs minimal change. Negative feedback clusters on exactly two things: solving the problem but not validating edge cases, and rewriting large portions of the code when the scale-up lands.
Yes — multi-tenancy is the through-line of everything.
Tenant-scoped rate limiting, quota enforcement and observable metrics are treated as non-negotiable — noisy-neighbour degradation is the failure they care about. Distributed architectures, caching, messaging queues, consistency models.
The one least served by memorising a catalogue — and the only one scoring thread safety first-class.
OA (24-hour window) → DSA round → LLD round → System Design + Hiring Manager.
Yes — explicitly Java, and explicitly concurrent.
A separate 60-minute machine-coding round based on Java, multithreading, concurrency and thread safety. Thread safety is a scored dimension here, not a follow-up question. The July 2026 Notification Service round was described as tough and deliberately ambiguous — significant time went just to understanding the problem.
Yes — round 3, after DSA and LLD.
Caching, sharding, consensus, rate limiting, plus the Salesforce-specific layer: multi-tenant data partitioning, org-level limits, metadata extensibility, zero-downtime deployment.
Strongest OOD bias of the big-tech names, and the AA round can sink you on its own.
Online Assessment → phone screen → onsite of 3–4 rounds, culminating in As Appropriate (AA). One reported onsite was two system design plus two technical.
Yes — and Microsoft is known for a strong OOD focus compared to other FAANG.
Coding rounds lean heavily on trees, graphs, strings and dynamic programming; design appears alongside rather than instead.
Yes — full system design rounds are expected from SDE-2 upward, and the weight increases with level.
Reasoning about reliability at scale. Panels balance coding rigour against collaboration signals.
Your best-matched target. Their system design round is the work you have actually done.
OA (CodeSignal or HackerRank, 70–90 min, 4 problems — most report needing 3 of 4 solved) → recruiter screen → DSA and fundamentals → system design (45 min, usually senior only) → behavioural on STAR. Three to eight weeks.
Not a dedicated machine-coding round. OOP fundamentals are probed inside the technical round, alongside OS and networking.
The technical round covers data structures, algorithms and core CS concepts — OOP, operating systems, networking — rather than a build-it-in-60-minutes exercise.
Yes — 45 minutes, and heavily emphasised.
Low-latency processing, fault tolerance, strict idempotency, high availability. Cache invalidation, sharding, replication, and SQL versus NoSQL for transactional consistency. They want you to think about reliability and edge cases deeply.
Both design rounds, plus the highest behavioural bar on this list.
Loop rounds are non-elimination. Each is 45 minutes of technical (DSA, LLD or HLD) plus 15 minutes of Leadership Principles. Most L5 loops carry one LLD and one HLD round; some teams combine them into a single 90-minute session. A Bar Raiser sits in the loop.
Yes — a standard part of the L5 loop.
Parking lot and elevator are among the three most commonly reported OOP prompts here — this is the one target where the classics really are the classics.
Yes.
Scaling storage and handling concurrent read/write at volume.
No design round at all. Spend nothing here on LLD or HLD prep.
Google Hiring Assessment → recruiter screen → one technical phone screen (45–60 min coding) → onsite of 2–3 coding rounds plus one behavioural (Googliness). Hiring committees weight coding most heavily.
No. Google reserves design for L5 and above.
No dedicated round at L4. Design concepts may surface inside a coding round; some variation by team exists.
A real target. The round shape is unlike anything else on this list, and it changed in April 2026.
Product Architecture, 45 minutes — define requirements, design the APIs, draw a high-level diagram.
Not a machine-coding round. No requirement to produce running code.
Yes, in the Product Architecture form.
Product mindset weighted above raw backend scalability: define requirements, design the APIs, draw the diagram. Scope varies by interviewer — some narrow it to API design or low-level component design.
The densest concentration of the machine-coding format, and the most drillable.
Flipkart, Swiggy, Cred, Razorpay, PhonePe, Ola, Groww, Udaan. Machine coding rounds run 90–120 minutes and are used from SDE-1 all the way through SDE-4.
Yes — this is where the round is most standardised.
Requirements are deliberately vaguer in India and Asia than in Western loops. You are expected to ask grounding questions, extract constraints and drive the solution yourself. Ninety to 120 minutes means finishing is expected, not optional.
Varies by company and level.
Where most of the new headcount is. Your DSA and LLD work transfers whole — this is additive, not a fork.
Recruiter and hiring-manager screen → one or two technical phone screens → sometimes a paid work trial → virtual onsite of four to five rounds. Six things get probed: ML/LLM fundamentals, applied ML, LLM/RAG engineering, coding, AI system design, behavioural.
Usually folded into the coding round rather than run separately.
Many loops KEEP a classic coding round — do not neglect data structures. Anthropic, OpenAI, Scale and Sierra ask their AI-infrastructure questions inside SWE and ML rounds, so the engineering fundamentals and the AI topics are prepared together, not instead of each other.
Yes — an open-ended 35–60 minute AI system design conversation.
Seven pillars tested as a set: RAG, fine-tuning, AI caching, agentic evals, AI observability, user feedback loops, AI guardrails. Almost every 2026 loop has at least one deep RAG round, and agents get a full round of their own. Questions are production-failure shaped, never definitional — they want to know what breaks and what you have shipped.
Nothing here rewards naming a pattern. Design and working code are 55% between them.
| Dimension | Weight |
|---|---|
| Code design & structure — class organisation, naming, separation of concerns | 30% |
| Working functionality — it compiles, it runs, it is correct | 25% |
| Design patterns & SOLID | 20% |
| Extensibility — add a feature without modifying existing code | 15% |
| Edge cases & error handling | 10% |
The code review is 30–50% of the evaluation. The code alone does not carry the round — interviewers ask why this pattern, how would you add X, what changes under concurrency, and they score the answers. A runnable driver main is non-negotiable; interviewers run it first.
| Phase | Minutes |
|---|---|
| Requirement extraction & design | 10–15 |
| Core models / entities | 15–20 |
| Business logic | 25–30 |
| Driver code | 10 |
| Edge cases & polish | 5–10 |
The classics have not gone away — but the 2025–26 direction of travel is toward ambiguous, product-shaped prompts. Memorising the catalogue is worth less than it was; handling an underspecified prompt is worth more.
Parking Lot · Snake and Ladder · Vending Machine · Shopping Cart
Elevator · Chess · Restaurant / Food Ordering · Movie Ticket Booking
Rate Limiter · Notification Service
Splitwise-style expense splitter · Ride-matching / Uber-lite · LRU cache with custom eviction · Food-delivery order tracking · In-memory file system · Cost Explorer · Middleware Router · Rating System · Customer Satisfaction system
LLD round landscape research (2026-08-18, this repo) · Uber SDE-2 L4 interview experience · Uber system design guide 2026 · Atlassian engineering interview guide · Atlassian system design guide 2026 · Salesforce SMTS interview experience · Salesforce MTS March 2026 · Microsoft India interview guide 2026 · Visa software engineer guide 2026 · Google L4 interview guide