[ ]productinterview
Run a free mock

system design · hard

"Design TikTok's system" PM interview answer

Design TikTok's system.

Updated Jun 2026 Calibrated to the strong-hire bar

This question tests whether you understand what makes TikTok’s product work, not whether you can describe a video transcoding pipeline. A weak answer lists infrastructure components. A strong one traces every architectural decision back to a product outcome: why completion rate is the objective function, why cold-start is a product viability question rather than a machine learning footnote, and why the content moderation tradeoff is fundamentally about creator trust versus ad revenue.

Scope it before you design

Open with a 60-second constraint: “I’ll focus on the FYP recommendation and content pipeline, the core loop that creates TikTok’s value for both creators and consumers. I’ll treat upload, discovery, and playback as the three subsystems, and flag where I’d make different choices at MVP versus scale. I’ll skip Live, DMs, and the creator monetization dashboard unless you want to go there.”

Then name the north star before drawing any boxes: the system’s job is to maximize the probability that a given user watches the next video to completion and returns the next day. Every architectural choice flows from that. An interviewer who hears you say that knows you understand TikTok’s product; an interviewer who hears you start with “so first we need a CDN” does not.

Scale context to anchor the conversation: TikTok processes roughly 34 million video uploads per day. That number shapes every non-functional requirement you’ll cite.

The three-stage FYP pipeline

The FYP is not a single model. It is a three-stage funnel, and the PM-relevant detail is what each stage optimizes for and what it costs to get wrong.

Candidate retrieval. The system pulls roughly 500 candidate videos from a corpus of billions using vector similarity search (Faiss or Milvus in practice). Vectors are content embeddings generated at upload time and stored in a vector database. This is the architectural choice that enables TikTok’s cold-start behavior: because retrieval is based on content similarity rather than engagement history, a video with zero plays can still reach users with matching interest embeddings. That is a deliberate product decision to keep creator entry barriers low, and it is what differentiated TikTok from YouTube in its early years.

Ranking. The 500 candidates are scored by a model trained with completion rate as the primary label, not clicks or likes. Supporting signals: user interaction history, video features (audio track, transcript, hashtags), and context signals like time of day and device type. The 2026 signal hierarchy, which interviewers from established creators now expect candidates to know: completion rate and re-watch loops are tier 1; shares to DMs and saves are tier 2; likes are tier 3. This ordering is a product choice, not an ML default. It rewards genuine interest over performative engagement, and it shapes creator behavior because creators optimize for what gets amplified.

One 2026-specific change to name explicitly: as of mid-2025, TikTok tests new videos with the creator’s existing followers first before opening non-follower distribution. This changed the cold-start dynamic. New creators with small followings now face a higher bar before reaching the FYP at scale; established creators get a follower-first signal that more accurately predicts whether a video earns broader distribution. The algorithmic amplification threshold for completion rate moved to approximately 70%, up from roughly 50% in 2024. An interviewer who works in or near the creator economy will notice if you cite the old number.

Re-ranking. The top-ranked videos go through a final layer that injects diversity (not ten consecutive videos on the same topic), applies policy filters and moderation holds, and runs creator freshness boosts for accounts that have not appeared in a user’s recent feed. This is also where the creator-consumer flywheel tension is resolved: a creator’s niche audience might want dense topical content, but FYP’s diversity requirements mean the system will occasionally interrupt that to keep session variety high. Naming this tension is a strong signal that you understand both sides of the two-sided marketplace.

Cold-start is a product viability question

Most candidates design for a user with 90 days of engagement history. Cold-start is harder and more strategically important on both sides of the market.

For new users: TikTok collects fast signals during onboarding (topic selection, the first few swipes) and uses content embeddings to surface relevant candidates immediately, without needing engagement history. The risk is false confidence: a user who swipes through five cooking videos in onboarding will get a cooking-heavy feed that may not reflect their actual breadth of interest. Recommend a diversity injection during the first session specifically to sample across categories.

For new creators: the content-embedding retrieval at upload time means a well-made video can reach matched audiences on day one. This is what enables the asymmetric creator economy TikTok built: a single video can generate millions of views before the creator has any followers. That is only possible because the system was explicitly designed to not require engagement history for discovery. Name this as a product decision, not an ML architecture detail.

One PM-relevant constraint interviewers sometimes probe: niche consistency. Creators posting across three or more unrelated topic areas see roughly 45% lower reach than creators who maintain topical focus. The system rewards topical authority as a proxy for content quality signal reliability. This is a product-designed behavior, and it is worth naming because it surfaces a real creator tension: breadth versus reach.

Content moderation as a PM judgment call

The moderation layer has two failure modes with very different costs. False positives (removing good content) erode creator trust and cause creator churn, which threatens content supply. False negatives (allowing harmful content) create brand safety risks that reduce advertiser willingness to pay, which threatens revenue.

The PM judgment call is where to set the operating point given your current business stage. Early: optimize for creator trust, accept a higher false negative rate, and invest in rapid human review for appeals. At scale with significant ad revenue: shift toward a lower false negative rate, invest in appeal turnaround time as the mechanism to manage creator trust, and be explicit about the category-specific thresholds (political content, health misinformation, and child safety each warrant different operating points).

What you should not do is present moderation as a single accuracy metric to optimize. It is a multi-objective problem with different stakeholder costs on each side of each error.

Data localization as a real design constraint

Post-2024 regulatory pressure made multi-region data architecture a genuine PM constraint rather than a theoretical one. US TikTok data is now stored on Oracle Cloud in the United States under Project Texas compliance requirements. The PM-level implication: content moderation, FYP serving, and user data storage cannot all be treated as globally unified systems. You need regional data boundaries in the architecture, with cross-region model training governed by data residency rules. Naming this shows you understand TikTok’s real operating environment, not a generic “design a video app” scenario.

Structure a strong answer

strong

"I'll scope this to the FYP recommendation and content pipeline: upload, discovery, and playback. The north star is maximizing the probability that a user watches the next video to completion and returns tomorrow. Every choice I make flows from that.

The FYP is a three-stage pipeline. Candidate retrieval pulls roughly 500 candidates from billions using vector similarity against content embeddings generated at upload time. This is what makes TikTok's cold-start work: retrieval is content-based, not engagement-based, so a new creator with zero followers can reach matched audiences on day one if the content quality is there. That is a deliberate product decision to keep creator entry barriers low.

Ranking scores those candidates with a model trained on completion rate as the primary label. Not clicks, not likes. The 2026 signal hierarchy: completion rate and re-watch loops are tier 1; shares to DMs and saves are tier 2; likes are tier 3. This ordering reflects what signals predict genuine interest versus performative engagement. One 2026 change worth naming: new videos are now tested with the creator's existing followers before opening to non-follower distribution. The completion rate threshold for algorithmic amplification is now approximately 70%, up from 50% in 2024.

Re-ranking injects diversity, applies policy filters, and resolves the creator-consumer flywheel tension: a creator's niche audience wants topical depth, but FYP's session diversity requirements interrupt that. The re-ranking layer is where that gets resolved.

On cold-start for new users: I'd recommend diversity injection during the first session specifically, because onboarding signals are noisy and a narrow early feed can trap users in a topic they don't actually want.

On content moderation: this is a PM judgment call between false positive rate, which causes creator churn, and false negative rate, which causes brand safety risk and advertiser revenue loss. I'd set different operating points by content category and stage of business, not a single accuracy target.

On data localization: US data is now stored separately under Project Texas. The architecture needs regional data boundaries that affect how the FYP model is trained and served across regions.

The metrics I'd close with: 7-day retention for new users, D7 re-upload rate for new creators, and average videos watched per session. These three validate that the system is working for both sides of the marketplace simultaneously."

weak

"TikTok needs a CDN, a recommendation engine, a video storage system, and a moderation layer. I'd measure success by total videos uploaded and DAU." This fails at every level a PM is evaluated on. It lists components without connecting any of them to outcomes. It picks DAU as a north star metric without noting that TikTok's differentiation comes from completion rate and return rate; DAU without session depth is a misleading signal. It treats the FYP as a black box, which means the follow-up question ("what signals does the model use?") will end the answer. It ignores cold-start entirely, which signals the candidate has not thought about the system's hardest and most strategically important edge cases. And it buries five minutes of the interview in infrastructure pattern-matching that an interviewer at TikTok does not need to hear from a PM candidate.

The 2026 reframe

In 2026, “design TikTok’s system” is really asking: how do you build a system where viability (advertiser revenue requiring brand-safe content at scale) and lovability (a feed that feels personally curated rather than algorithmically obvious) are both achieved through the same infrastructure decisions?

The completion rate optimization is the answer. It is the signal that aligns what users genuinely love, what creators are rewarded to make, and what advertisers are willing to pay for. The system design question is fundamentally about proving you understand that alignment. Feasibility is not the interesting constraint; serving a billion users a personalized feed under 200ms is solved infrastructure. The PM-interesting questions are whether the cold-start solution keeps creator entry barriers low enough to sustain content supply, whether the moderation system can be calibrated without becoming a moat for established creators, and whether the diversity injection in re-ranking can resolve the niche-consistency tension without breaking the creator’s sense of topical authority. Those are viable and lovable questions wearing a system design costume.

Asked at

Related