System design — the capstone walkthrough

The interviewer says "design Instagram Reels ranking" or "design Google Sponsored Search." This is the answer, with every previous lesson slotted into its place. The latency-budget simulator at the bottom is the arithmetic you should be ready to do out loud.

The interview has a predictable rhythm — ride it

Most system-design interviews follow the same arc, regardless of which product the interviewer names. Senior candidates don't fight it; they ride it, because the rhythm is also a natural way to actually design a system.

1. Clarify scope. Scale, latency, single- vs multi-task, cold start, what counts as success. Two minutes, five questions.
2. Define metrics. Top-line online metric, offline proxy, guardrails. One minute.
3. Draw the funnel. Retrieve → rank → re-rank, with K₁, K₂, K_shown and a per-stage latency budget. Two minutes. This is the diagram you must be able to draw in 30 seconds.
4. Zoom in where steered. The interviewer will pick a stage — cold start, debiasing, the auction, calibration drift. The remaining 25 minutes lives here.
5. Failure modes and operations. Train/serve skew, calibration monitoring, A/B test power, on-call playbook. Five minutes, often near the end.

Worked example 1 — "Design Instagram Reels Home ranking"

Step 1 · Clarify

Five questions, asked out loud:

• Scale. Roughly a billion monthly users, on the order of 10⁹ active short videos, ~10² impressions per session.
• Latency. End-to-end per-impression on the order of 50–100 ms; the ML budget is maybe 40 ms inside that.
• Cold start. Both. New users (no history) and new videos (just uploaded) both matter — Reels grows by both.
• Objective. Multi-task. Watch time, like, share, retain. Not just CTR — clickbait is a known failure mode for video.
• Constraints. Diversity (no five basketball clips in a row), creator caps, ad slots inserted, policy filters.

Step 2 · Metrics

Step 3 · Draw the funnel

Step 4 · Each stage in one breath

Retrieve (lessons 2–3). Multiple sources because no single retrieval signal covers the space. A collaborative two-tower (user_id × video_id, trained on watch/like history) for behavioral affinity. A content two-tower (text/audio/video embedding × user-history embedding) for new videos with no engagement signal. A recent-network source pulling videos posted by accounts the user follows. A trending baseline. An exploration source — a small budget for items the system is uncertain about. Each source has its own index (HNSW for the dense towers, inverted index for trending). Union, dedupe, hand the top K₁ to the ranker.

Rank (lessons 4–5). A DLRM-class network. Sparse features: user_id, video_id, creator_id, hashtags, audio_id. Dense features: video age, duration, prior engagement rates, geo. User history: last 100 actions, pooled (mean or small transformer). Multi-task heads share the trunk: sigmoid heads per binary signal (P(watch ≥ 5s), P(like), P(skip), P(share)) plus a regression head for E[watchtime]. Final score is a hand-tuned linear combination — policy weights, not learned end-to-end.

Re-rank. Listwise constraints the per-item ranker can't see. Creator caps in the shown window, topic diversity, ads inserted at policy positions, one exploration slot in the first 10, policy filters.

Step 5 · Train and eval

Pointwise BCE per binary head + MSE / Huber on log-watchtime. Daily retrain from yesterday's impressions joined with engagement labels (24-hour attribution window). Position-as-feature at train time only, set to a constant (e.g., position 0 or 1) at inference (lesson 7). Calibration re-fit hourly because the daily model drifts. Offline: NDCG and AUC per head, calibration error on a held-out hour. Online: A/B on session watch-time and DAU as primary; CUPED variance reduction (lesson 8). Long-term holdout cohort to catch retention regressions a 2-week A/B can't see.

Step 6 · Cold start

New users: skip the collaborative tower, retrieve from content-only and trending until the first ~10 sessions, then mix in the collaborative source. New videos: compute content embeddings at upload time, push into the ANN index immediately; the exploration source guarantees first impressions so the video has behavioral signal by the time the collaborative retriever needs it.

Worked example 2 — "Design Google Sponsored Search ads"

Step 1 · Clarify

• Scale. Billions of queries per day; a long-tail head distribution — the top 1% of queries get most of the traffic.
• Latency. ~100 ms end-to-end for the ad system, parallel with organic search.
• Surfaces. Multi-slot — top mainline, bottom mainline, side. Each has different reserves.
• Objective. Three-way: revenue, user click quality (don't show bad ads), advertiser efficiency (advertisers reaching their CPA targets).

Step 2 · The funnel — five stages, not three

Step 3 · Each stage in one breath

Match is ads-specific. Three modes. Exact match: query string == bid keyword. Phrase match: query contains the bid keyword. Broad match: semantic — a query-tower and keyword-tower (two-tower again) decide whether the query is close enough to the advertiser's keyword. Broad match is itself a recall problem and uses retrieval-style ML.

Pre-rank. A cheap CTR model (linear or shallow MLP) scores the eligible 10²–10⁴ ads. Keep the top ~100.

Rank. Full DLRM. Sparse features: query × ad cross, advertiser_id, keyword_id, vertical. Two heads: P(click) and P(conversion | click). Calibration is mandatory — the auction multiplies pCTR by bid, so a calibration error is a revenue or efficiency error of the same size. Position-feature at train time, set to a constant (e.g., position 0 or 1) at inference.

Auction (lesson 9). Rank-score = bid · pCTR · quality_score. GSP or VCG-style depending on surface. Reserve prices per query.

Pricing. Under GSP, the price for position k is score_{k+1} / pCTR_k — the minimum that would have kept you in your slot. VCG pays the externality you impose.

Autobidder (lesson 10). In parallel on the advertiser side, a separate system per advertiser account sets the per-impression bid to hit a target CPA or ROAS, using the same calibrated pCTR/pCVR signals.

Step 4 · Train and eval

Pointwise BCE for the click head; BCE for the conversion head trained only on clicked impressions (beware selection bias). IPS-debias on position. Daily retrain. Calibration monitoring is the on-call alarm — drift beyond threshold on a major vertical pages someone. Revenue eval: A/B by advertiser segment, advertiser CPA as secondary, user-side click quality (search abandonment, refinement) as a guardrail.

Operational concerns every senior interviewer will probe

The five most common follow-up directions

Whichever product the interviewer names, one of these five follow-ups arrives:

Interactive · latency-budget timing simulator

Lesson 1 said the budget arithmetic is the load-bearing senior signal. This widget makes it concrete. You allocate milliseconds across four pipeline stages out of a 100 ms end-to-end budget; under each stage the widget tells you how many items that budget can score given a representative per-item cost. The diagnosis at the bottom names the bottleneck — the stage that constrains end-to-end K — and tells you whether the funnel's K-values are internally consistent.

What separates senior from junior, distilled

Interview prompts you should be ready for

That's the full search/ads/recsys series. The interactive widgets in lessons 1, 5, 7, 8, 9, and this one are the ones worth replaying before an interview — they exercise the arithmetic you'll need to do out loud. Good luck.