How would you design a news feed system?

// News feed: capacity, ranking candidates, and why ranked feeds paginate from a snapshot.
const fmt = (n) => n.toLocaleString("en-US");

console.log("== 1. Capacity (assumptions: 300M DAU, 10 feed opens/day, peak = 3x avg, 500 candidates/request) ==");
const dau = 300_000_000, opens = 10, cands = 500;
const perDay = dau * opens;
const avg = Math.floor(perDay / 86400);
const peak = avg * 3;
console.log("feed requests/day   " + fmt(perDay));
console.log("avg requests/s      " + fmt(avg));
console.log("peak requests/s     " + fmt(peak));
console.log("candidates scored/s " + fmt(peak * cands) + " at peak (one heavy model on all of them is too costly)");
console.log("stage 2 scores/s    " + fmt(peak * 50) + " if a light stage 1 keeps the top 50");
console.log("feed id cache       " + fmt(dau * cands * 8 / 1e9) + " GB (500 post ids x 8 bytes per user)");

console.log("\n== 2. Ranking: score = affinity x type weight x 10000 / (age_min + 120) + likes ==");
const affinity = { mom: 90, ana: 60, bob: 30, brand: 10 };
const typeW = { photo: 3, video: 3, link: 1, text: 2 };
// [id, author, type, age in minutes at session start, likes]
const posts = [
  ["p1", "brand", "video", 5, 400], ["p2", "bob", "text", 10, 2], ["p3", "ana", "photo", 30, 40],
  ["p4", "mom", "photo", 240, 15], ["p5", "bob", "link", 45, 1], ["p6", "ana", "text", 60, 8],
  ["p7", "brand", "link", 90, 900], ["p8", "mom", "text", 600, 3], ["p9", "bob", "photo", 120, 12],
  ["p10", "ana", "video", 300, 25],
].map(([id, author, type, age, likes]) => ({ id, author, type, age, likes }));
const score = (p) => Math.floor(affinity[p.author] * typeW[p.type] * 10000 / (p.age + 120)) + p.likes;
for (const p of posts) p.score = score(p);
const byTime = [...posts].sort((a, b) => a.age - b.age || (a.id < b.id ? -1 : 1));
const byScore = [...posts].sort((a, b) => b.score - a.score || (a.id < b.id ? -1 : 1));
console.log("chronological: " + byTime.map((p) => p.id).join(" "));
console.log("ranked:        " + byScore.map((p) => p.id + "(" + p.score + ")").join(" "));

console.log("\n== 3. Page 2 is requested 30 min later: ages grow, p9 gains 1500 likes, 2 new posts (page size 4) ==");
const fresh = [{ id: "n1", author: "ana", type: "photo", age: 0, likes: 0 }, { id: "n2", author: "mom", type: "photo", age: 0, likes: 0 }];
const later = posts.map((p) => ({ ...p, age: p.age + 30, likes: p.likes + (p.id === "p9" ? 1500 : 0) })).concat(fresh);
for (const p of later) p.score = score(p);
const rankedLater = [...later].sort((a, b) => b.score - a.score || (a.id < b.id ? -1 : 1));
console.log("re-ranked now: " + rankedLater.map((p) => p.id + "(" + p.score + ")").join(" "));
const page1 = byScore.slice(0, 4).map((p) => p.id);
const report = (name, p2, unreachable) => {
  const dup = p2.filter((id) => page1.includes(id));
  console.log(name.padEnd(24) + "page2=[" + p2.join(" ") + "] duplicates=" + dup.length + " never shown=" + (unreachable.join(" ") || "none"));
};
console.log("page1 (all methods)     [" + page1.join(" ") + "]");
// offset pagination: re-run the ranking, skip the first 4
const offsetP2 = rankedLater.slice(4, 8).map((p) => p.id);
report("offset re-query", offsetP2, rankedLater.slice(0, 4).filter((p) => !page1.includes(p.id) && !p.id.startsWith("n")).map((p) => p.id));
// keyset cursor on (score, id) of the last item on page 1
const cur = byScore[3];
const below = (p) => p.score < cur.score || (p.score === cur.score && p.id > cur.id);
const keyP2 = rankedLater.filter(below).slice(0, 4).map((p) => p.id);
report("keyset cursor on score", keyP2, rankedLater.filter((p) => !below(p) && !page1.includes(p.id) && !p.id.startsWith("n")).map((p) => p.id));
// snapshot: rank once at session start, store the ordered ids, page by position
const snapshot = byScore.map((p) => p.id);
report("session snapshot", snapshot.slice(4, 8), []);
console.log("snapshot banner: " + fresh.length + " new posts (shown at top on refresh)");

console.log("\n== 4. Next session: seen posts filtered, new posts ranked in ==");
const seen = new Set(snapshot.slice(0, 8));
const next = rankedLater.filter((p) => !seen.has(p.id));
console.log("seen=" + seen.size + " next feed: " + next.map((p) => p.id + "(" + p.score + ")").join(" "));