What TikTok Tests in System Design

System design carries real weight in the TikTok loop, and unlike coding it tends to decide your level rather than the yes-or-no. The round runs 45 to 60 minutes and shows up from mid-level (2-1) upward, with more of it as you go senior. Some senior and lead loops report two design rounds instead of one: a product-space prompt in one, and a deep dive into a system from your own resume in the other. Remember the whole loop is ByteDance's, so recruiters, levels, and the products in the prompt all sit under ByteDance even when the role says TikTok. Coding still gates the hire because you have to write running, tested code live. Design sets the rung. That is where a senior offer is won.

The TikTok angle

The real-time recommendation loop behind the For You feed is TikTok's signature design, and it is the one thing that separates a TikTok answer from every other recsys answer. The For You feed, the endless personalized video stream most people call the FYP, is powered by a system that learns online, within minutes, from the watch-time and skips it just collected. ByteDance calls it Monolith, and it is genuinely published: an arXiv paper and an open-sourced framework. Everyone can draw a candidate-generation to ranking funnel. To sound like a TikTok engineer, open with the fast feedback loop. Engagement events flow into a log. A joiner pairs each signal with the features that served it, the training side updates, and fresh parameters reach serving in about a minute. Draw a static two-tower model refreshed every few hours instead and you have drawn the Pinterest or Snap answer, which is close to the opposite of what makes TikTok fast.

The round, and which prompt you get

Two flavors of design round show up in candidate reports, and you can get either or both. The first is a product-space prompt in TikTok's own surface: design TikTok, design the For You feed, design real-time messaging, design the upload and delivery pipeline, design TikTok LIVE. The second is a project deep dive. The interviewer takes a system off your resume and pushes on scale, bottlenecks, and trade-offs until it either holds up at TikTok's numbers or does not. Prep both. The grid below is the shape of the round. ByteDance publishes no question list, so read it as a steer toward where prep pays, not a fixed bank.

• One round, sometimes two: 45 to 60 min. Expected from mid-level (2-1) up, weighted more heavily at senior (2-2). Several senior and lead loops report two design rounds, sometimes labeled cross-functional: one product prompt, one project deep dive. New-grad and frontend loops may fold design into a knowledge round instead.
• Product prompt: TikTok's own surface. Design TikTok is the signature, usually scoped to the For You feed plus video serving. Real-time messaging and an online-offline presence provider are the most-attested prompts after it. Upload and delivery, TikTok LIVE, notifications, and ad ranking round out the set.
• Project deep dive: Your own system at scale. Not a design-X prompt at all. The interviewer drives into a system you built and asks you to defend it. One report described a rejection after a candidate answered with a 50-per-second design when the follow-up wanted millions per second. Bring a system you can push to TikTok scale.
• Team-dependent: recsys, T&S, ads, infra. Hiring is decentralized, so the prompt is flavored by the org you interview with. The recommendation org leans into the feed loop, ads teams into auctions, trust-and-safety toward practical design. Confirm your team and its likely surface on the recruiter call.

Two habits carry across both flavors. Interviewers ask for large consumer-scale reasoning, so a single-box answer that never reaches sharding, fan-out, or failure will stall. And because the same senior deciders who set the bar often sit in China and hold the real call, the round can be blunt about whether your design survives at scale. That project-deep-dive rejection is the tell: the number that matters is the one you can defend, not the one that sounds impressive.

The TikTok angle

Come with one system from your own work that you can push from its real numbers to TikTok's numbers on the spot. Do not treat it as the easy one. The project deep dive is not a softer round. Candidates describe it starting friendly, a walk through something you built. Then it turns into a scale interrogation: what breaks at ten times the load, at a hundred times, where the bottleneck moves, what you would shard first. One report ended in a rejection when a candidate defended a design at fifty requests per second and the interviewer wanted millions. Pick a system where you know the throughput, the storage footprint, and the failure modes cold, then rehearse walking it up the orders of magnitude. The recommendation, ads, trust-and-safety, and infra orgs each hire on their own bar, so match the system you bring to the team you are interviewing with.

What is published, and what is reconstructed

Honesty up front, because it changes how you use this section. TikTok's feed is unusually well documented for a recsys, and the rest of the stack is not. ByteDance published Monolith, its real-time recommendation training and serving system, as an arXiv paper (2209.07663) and open-sourced the framework. That gives you real, citable design to reason from. The table is collisionless, built on Cuckoo hashing, with frequency filtering and expirable embeddings that keep it affordable. A training parameter server and a serving parameter server sync sparse embeddings about once a minute. And the design trades some reliability for freshness on purpose.

The rest of the stack is different. The upload and transcoding pipeline, the CDN delivery path, TikTok LIVE, and the notification fan-out are not published. ByteDance has not published deep dives on those, so this section builds them as the standard large-scale user-generated-video design at TikTok's scale, with the TikTok-specific couplings called out. Treat those as well-founded engineering, not as a map of ByteDance's servers.

The TikTok angle

Ground the feed answer in Monolith's published ideas, and reason from first principles everywhere else. In the room, naming the collisionless Cuckoo embedding table, the minute-level sparse sync, and the freshness-over-reliability trade-off shows you have read the real design, because those are published. What a strong answer avoids is claiming ByteDance's current production topology does exactly some specific thing. The 2022 paper describes a framework, not today's live system, and an interviewer who works on it will know the difference. For upload, delivery, and LIVE, say plainly that you are applying standard video-infrastructure design at TikTok scale. Then layer on the couplings that are genuinely TikTok's. A new upload has to be transcoded, moderated, and embedded fast enough to enter the freshness loop within minutes.

The shapes the round lands on

Group the reported prompts by the underlying shape rather than the wording, and a small set covers almost everything. The chart orders them by how often candidates on the relevant track meet each one.

Each shape forces its own hard problem. The recommendation feed is about the freshness loop and the multi-objective ranking that follows it. Real-time messaging and presence is about long-lived connections, an offline message store, and status that expires when a heartbeat stops. It has no standalone lesson here. The notification design works the same long-lived-connection and per-room fan-out mechanics, so use it as your worked reference. Video upload and delivery is about chunked resumable uploads over flaky mobile networks and a transcode fan-out that races to make a clip watchable. Live streaming pairs a low-latency video path with a chat and gifting path that has to stay correct about money. Notifications is about fan-out that does not collapse under a creator with fifty million followers. Ad ranking is a targeting-to-auction funnel on the same sparse-embedding machinery as the feed. Same machine, different objective.

What the round rewards

Five moves carry any of these prompts without thrashing. Pin the objective and the number you are optimizing first. For the feed that is watch-time and feed latency under a couple hundred milliseconds. For upload it is time from post to watchable, and for LIVE it is glass-to-glass delay. Lay the request or data path end to end before you zoom in, so the interviewer sees the whole machine. Name the real bottleneck out loud. Each shape has one the generic version glosses over: the freshness sync for the feed, the transcode-and-moderate race for upload, the money path for LIVE, celebrity fan-out for notifications. Then tell the scaling and failure story, and close on cost. Go deep on one stage and keep the rest brisk. The back half is what separates the strong answers.

Below is the reference loop to carry for any feed answer. It is drawn from Monolith's published design, and the same instinct, collect signals, learn fast, serve fresh, runs through most TikTok recsys questions.

Read the loop in chips. A Viewer App pushes engagement into an Event Log, a Joiner + Sampling step attaches each label to the features that served it, and the Training Param Server updates the Cuckoo embedding table. Fresh parameters reach the Serving Param Server about once a minute, and a feed request runs through Candidate Generation, Multi-stage Ranking, and Diversity Re-rank into a Prefetched Feed. The instinct to carry: freshness is the wedge, watch-time outranks likes, and the sparse-versus-dense sync cadence is the detail that separates a deep answer.

The worked designs

Five designs follow, one per distinct shape, ordered so the signature comes first. Start with the For You feed: it is the fullest version of the loop above and the design no other guide in this catalog teaches the same way.

Each design assumes the fundamentals from the System Design course: the scaling playbook for read and write paths, and the toolbox for when to reach for a log, a cache, a queue, or a CDN. These pages stay on the TikTok-flavored design and leave the general mechanics to those.