Lead: the prompts look familiar; the pass bar does not
The prompts still sound like Instagram, a rate limiter, a GPU batch API. What moved in 2026 is who still gets an offer after a “pass.” More people clear the technical floor than there are seats. Interviewers split the pool on cost, how the thing dies, and whether you would notice at 3 a.m.
This page is the conversation shape: what changed, which prompt families show up, a five-step clock, week plans. Then a human on Find a partner. If CAP vs quorum still stalls, close it on express quizzes first. Building blocks live in the 12-concepts sheet; titles to open after a miss are on the 2026 books list.
Disclaimer. Practice map, not a FAANG loop and not payroll advice. Company formats and “what candidates heard” are public-roundup order of magnitude. The gym does not grade sketches, does not host the call, and does not promise a rating.
Key takeaways
- Passing the old bar is often a “hire, not offer.” Stand out inside the pass pool: cost, failure, ops.
- LLM serving and GPU batching now show up in general SWE loops. Treat them as queues and constrained compute.
- Some rooms dropped the whiteboard. Ship a verbal v1, or a sequential build that must run before v2.
- Five steps, ~45 minutes: scope, sketch, deep-dive two pieces, scale/cost/ops, wrap.
- Memorized Twitter boxes fail. Compose blocks, then say them under a timer with a partner.
What the room is
Forty-five to sixty minutes. An open prompt. You pick components, data path, and the trade-off you will defend. There is no answer key. They are scoring how you think, not whether you named Kafka.
| Slice | What they actually watch |
|---|---|
| Clock | Usually 45–60 minutes; some compressed verbal slots |
| Surface | Whiteboard, shared board, or no diagram at all |
| Score | Thought process, trade-offs, failure, cost, how you would operate it |
| “Correct” | Several architectures can pass; a memorized blog diagram often fails |
What moved in 2026
The pass pool is crowded
Public hiring-loop reports from 2026 say the same scores that used to convert now stall at committee. That is not one company. Headcount is the bottleneck. You need a signal inside the people who already “passed”: ops, cost, a failure story that is not “add servers.”
LLM infra is no longer an ML-only round
Serving a model, a support bot on a third-party LLM, shipping weights over a thin link, allocating GPUs, putting a guardrail on an agent that can act — these now appear in general software loops. The patterns are still queues, batching, backpressure, and blast radius. Close a hole on deep quizzes if “async” is still a slogan.
The board is optional; sequential builds are not
Classic whiteboard remains the default. New shapes: a multi-part code build that opens up only once v1 runs; a purely verbal Amazon-style GenAI slot; a reverse “defend a system you shipped.” Practice a working v1 before decorating v2. Adjacent walkthrough: Stripe system design.
Cost and ops are scored, not bonus trivia
“We would scale out” is not an ending. Spot vs reserved, cache vs recompute, serverless that is cheap until it is not, how the on-call would know. Say monitoring before they ask. That is the 2026 split.
Prompt families (what people actually heard)
Not a homework list. Families so you can reuse blocks. On a mock, pick one row and stay inside it for 45 minutes.
| Family | Example shape |
|---|---|
| AI / LLM infra | Query path for an LLM; support bot on a vendor model; GPU batch inference; weight fan-out; agent with tool calls |
| Classic distributed | OTA updates; URL shortener; KV store; rate limiter; queue; crawler; typeahead |
| Social / real-time | Feed, DMs, notifications, short video |
| Media / delivery | VOD, recommendations, object storage / sync |
| Pipelines | Event ingest; metrics; crawl → index |
| Commerce | Tickets, hotels, rides, payments, PDP at spike |
| LLD / APIs | LRU cache; collab editor (OT vs CRDT); in-memory store |
How loops differ (order of magnitude)
Use this to pick the mock flavor, not to name-drop a product you have not used. Gym does not simulate these rooms. Company deep-dives on the blog: Amazon, DoorDash.
| Loop | Clock | What to practice |
|---|---|---|
| ~45m | Engine choice by workload; do not invent their internal names as “the answer” | |
| Meta | ~45m | Feeds, messaging, CDN, always scale in the first sentence |
| Amazon | 20–45m | Product-shaped prompts; behavioral woven in; sometimes no board |
| Apple | 45–60m | Privacy, auth, APIs, storage; sometimes two design-shaped rounds |
| Netflix | ~60m | Open discussion; recs and internal tools; diagrams optional |
| Airbnb | 45–60m | Vague prompt on purpose; failure and “how you think” |
| OpenAI / Anthropic | 45–60m | 10×/100×/1000×; GPU batch as constrained compute, not magic |
| Stripe | ~45m | Sequential parts; v1 must work |
One prompt, out loud, on a clock
Reading the table is not the drill. Already have an account? Log in and match.
Five-step clock (~45–60 minutes)
Steal the clock, not a script. If the interviewer hijacks step 3, let them. Adjacent Amazon-shaped timing: Amazon system design.
-
Clarify — 5–8 min
Functional slice (upload vs feed vs live). NFRs: availability, consistency, latency, cost — which wins. Users, RPS, size. Read vs write. Say what you are not building.
-
High-level — 8–10 min
APIs, services, stores, caches. One data-flow per core use case. REST / gRPC / WS because of a requirement, not habit.
-
Deep dive — 10–15 min
One or two components. Happy path, how it fails, what the client sees. Push 10×/100× if they ask. Pick an engine for this workload, not “SQL vs NoSQL” as a meme.
-
Scale, cost, ops — 5–8 min
Bottlenecks, SPOFs, reserved vs spot, cache vs recompute, serverless cliffs. Logs, metrics, traces, who pages at 3 a.m. — unprompted.
-
Wrap — 3–5 min
Restate trade-offs. What you would add with more time: migration, finer ACL, geo failover — not a second architecture.
Week plans on the gym (not a second reading list)
| When | Focus | On IT Career Gym |
|---|---|---|
| Weeks 1–2 | Compose LB, cache, store, queue, CDN, hash, replica, shard | Course + 12 concepts |
| Weeks 3–5 | 8–10 prompts, 45 min, out loud, even solo | Quizzes then a daily challenge |
| Weeks 6–7 | Partner mocks; deep-dive failure | Find a partner — you host |
| Week 8 | Company flavor from the table; one weak block | Deep quiz + a short write-up |
Two weeks only: compress fundamentals into three days, then company prompts, then at least two mocks. AI-first posting: add a week on batch inference, GPU pools, RAG as retrieval+eval, LLM rate limits — still queues, not a Kaggle medal. The common miss is memorization.
Blocks to compose (not reciting)
| Block | Say this on the mock |
|---|---|
| Load balancing | L4 vs L7 for this traffic, not “we put an LB” |
| Caching | Aside / through / behind; invalidation; when stale is dangerous |
| Stores | Name an engine that fits the access pattern |
| Queues | Stream vs job vs fan-out |
| CDN | Edge for media; origin for personalized |
| Hashing | Add/remove nodes without reshuffling the world |
| Replica vs shard | Durability/reads vs write scale — different jobs |
| CAP | On partition, which side you pick and why users can live with it |
| APIs | REST default; gRPC inside; WS if bidirectional is required |
| Observability | Logs, metrics, traces, alert — unprompted |
What the room rewards by level
| Level | Signal |
|---|---|
| Junior | A working sketch, tests, obvious failure named |
| Senior | Scale, reliability, maintainability; how today’s cut shapes next year |
| Staff | Sets scope, names other teams, picks the two deep-dives without being asked |
Harder bar, same blocks
Compose, fail out loud, cost and ops without a prompt, LLM infra as distributed systems if the posting ships AI. Adjacent: 12 concepts, books, pair mock, last week.
Adapted from a public 2026 prep roundup and rewritten for a practice gym. The gym is free to start. Book the mock; let the quiz fill the hole.
Questions people actually ask
One row open. Follow-ups after “how do I prep?”, not a glossary of CAP.
-
How long is the design loop?
Most loops are 45–60 minutes. Some Amazon GenAI slots compress design into a short verbal section. Netflix often sits at a full hour. The gym mock is a peer call you host — not that company’s timer.
Link to this answer -
Do I have to draw?
Not always. Some loops stay verbal. Stripe-style builds want a working v1 before v2. Practice stating the data flow without a diagram, then add boxes only if the partner asks.
Link to this answer -
Senior versus staff — what actually changes?
Senior still owns scale, reliability, and a product-shaped trade-off. Staff is expected to set scope, name cross-team cost, and pick the two components that matter. Gym does not assign those levels.
Link to this answer -
Do I need ML for a 2026 software loop?
Only if the posting ships AI. Then treat LLM serving as queues, batching, and constrained GPUs — not as a model zoo. Close the hole on quizzes, then say it on a mock.
Link to this answer -
How is this different from the 12-concepts article?
The 12-concepts page is the building-block cheat sheet. This page is the 45-minute conversation: bar, prompt families, timing, company formats.
Link to this answer -
How is this different from the 2026 books list?
Books are what to open after a failed follow-up. This guide is how the room runs. Read both; drill on the gym.
Link to this answer -
Does Gym run a FAANG design interview?
No. Match a human on Find a partner, you host the call. Timed quizzes catch the stack. The gym does not grade your sketch or promise an offer.
Link to this answer -
Where do I start this week?
Name one prompt, walk the five steps out loud, close a hole on an express quiz, then book a pair mock. Do not collect a second reading list.
Link to this answer
Run the clock. Fill the hole after.
Open matching. Keep quizzes for the stack. Already have an account? Log in.
Related: 12 system design concepts · 2026 books · Stripe · Amazon · Pair mock