Design Reviews & Ratings System

Problem Statement: Reviews & Ratings Platform

A reviews and ratings platform is the social-proof engine of a marketplace: it captures a shopper's 1-to-5 star rating and written experience for a product, then folds millions of those opinions into the single number and star histogram that sits beside the Add to Cart button. That number moves money. On a large marketplace a half-star lift on a product detail page (PDP) measurably raises conversion, so this is not a comment box bolted onto the catalog — it is a ranking, aggregation, and trust system the buying decision depends on.

At its core the platform does three things that pull in opposite directions. It must accept writes that are correct and abuse-resistant (the right person, who actually bought the item, leaving one honest review); it must serve reads that are fast and always-on (the PDP renders the aggregate in tens of milliseconds for every shopper, reviewer or not); and it must moderate content that may be spam, fake, or unsafe before it poisons the average. Reads outnumber writes by roughly 100 to 1, so the architecture is fundamentally a read-optimized fan-out of a comparatively rare, carefully guarded write.

The defining tension is consistency versus availability, decided per surface. An individual review row — its publish state, its edit history — is treated as CP: we would rather delay or reject a write than show two conflicting versions of someone's review. The PDP aggregate (average stars, count, histogram) is treated as AP with bounded staleness: a shopper seeing 4.3 stars that is forty seconds stale is fine, but a PDP that fails to render because the aggregate store is briefly unavailable is not. Naming which surface gets which guarantee, out loud, is the move that separates a senior answer from a junior one.

Scale frames every later decision. Assume 80 million monthly shoppers browsing 12 million SKUs, generating about 1.2 million new reviews per day against hundreds of thousands of PDP renders per second at peak. The hard cases are not the average product with forty reviews; they are the hot SKUs — a console launch, a viral product — where reads and writes concentrate on a single key and naive per-product locking or recomputation falls over.