System Design Interview
Layered moderation for a large community, inspired by Discord and Reddit: known-bad content matched by hash, a fast classifier that decides most posts in milliseconds, an LLM that reviews the uncertain ones against the written rules, and human moderators for the hardest cases and appeals. A teaching design.
See why running the biggest model on every post is too slow and too expensive.
What was missing: Nothing yet.
Why that's risky: Over a second per post and a large bill per post; when the LLM is slow, nobody can post.
What gets added: One server that sends every new post to a large LLM and only saves it once the model says it is fine.
Trade-offs: The most accurate single check, applied where it is least needed.
Decide most posts in milliseconds with a small model.
What was missing: Every post waited on an expensive LLM.
Why that's risky: Posts in the middle band are either let through or blocked by a guess.
What gets added: A small classifier on GPUs scores every post in about 15 ms. High scores are blocked, low scores go live, and every moderation decision is saved with its scores. Feeds read posts through a cache.
Trade-offs: The small model is fast and cheap but misses context, sarcasm and new slang.
Spend the expensive model only where it is needed.
What was missing: Borderline posts were decided by the small model's guess.
Why that's risky: A borderline post can be visible for a few minutes before the LLM hides it; hold uncertain posts from large audiences until decided.
What gets added: Posts in the uncertain band (about 10%) are queued. LLM workers judge each one against the written community rules, save the decision with the model's reasoning, and hide the post if needed.
Trade-offs: Accuracy where it matters, at a tenth of the cost of checking everything with the LLM.
Block re-posts instantly and route the hardest cases to people.
What was missing: Spam waves re-posted the same content thousands of times, and nobody handled what the LLM itself was unsure about or what users appealed.
Why that's risky: Moderators see disturbing content; limit exposure per shift and blur images by default.
What gets added: A cache of perceptual hashes of removed content, checked before anything else. Cases the LLM is unsure about, and appeals, go to a human review queue; assignment workers route each case to a moderator by language and topic, and the moderator tool records their decision.
Trade-offs: People are slow and expensive, so they only see what machines can't decide.
Survive spam floods and remove single points of failure.
What was missing: Bot floods could overwhelm posting, and the load balancer and databases were single points of failure.
Why that's risky: Strict limits on new accounts slow down genuine newcomers.
What gets added: Rate limits per account (stricter for new ones), a second load balancer, sharded post storage, replicated decisions, more classifier capacity, and monitoring of appeal reversals to catch a drifting model.
Trade-offs: Appeal reversals are the best signal that a model needs retraining.
You're in the middle of an interview session. Leaving now will end your current attempt.
Explore concept overviews, real-system examples, key tradeoffs, and interview talking points for each roadmap section.
You've conquered this phase. These are the skills you now own:
