Design Social Search - System Design | WinJob

Problem Statement: Social Search at Twitter Scale

Design platform-wide social search like Twitter/X, Facebook, or LinkedIn: users type a query and retrieve posts, people, hashtags, photos, and videos with typo tolerance, synonyms, and filters (date, media type, from-user). The product is not a standalone search engine—it is the discovery layer on top of an existing social graph and content graph.

Interviewers expect you to separate read-heavy query path (p95 < 200ms) from write-heavy indexing path (seconds of lag acceptable with clear SLA). You must explain how a new post becomes searchable without blocking the publish API.

Phase lens (understanding)

Frame the product as discovery on a social graph—not a standalone web search engine.

Operational detail (sec-01)

Document visibility rules early: blocked users, deleted posts, and NSFW tags must filter at index time and query time. Use monotonic version on post_id so out-of-order Kafka events cannot resurrect deleted content.

Failure modes unique to this step

Viral hashtag saturates one shard; suggest trie poisoned by bot queries; cross-region replication lag shows stale results after move—mitigate with rate limits, suggest sanitization, and X-Index-Lag-Sec response header.

Interview checkpoint

State one outage story for "Problem Statement: Social Search at Twitter Scale" (stale index, hot shard, typo spam) and the control that limits blast radius.

javaOne Dark Pro

1 public record SearchCursor(String searchAfterScore, String docId, long sessionTtlMs) {
2   public SearchCursor {
3     if (sessionTtlMs < 0) throw new IllegalArgumentException("ttl");
4   }
5 }

pythonOne Dark Pro

1 def blend_scores(post: float, people: float, tag: float, intent: str) -> float:
2     weights = {"posts": 1.0, "people": 0.9, "hashtags": 0.7}
3     base = post * weights["posts"]
4     if intent == "people":
5         base = max(base, people * 1.2)
6     return base + 0.15 * tag

typescriptOne Dark Pro

1 export function normalizeQuery(raw: string): string {
2   return raw.normalize("NFKC").trim().slice(0, 512).toLowerCase();
3 }

Why interviewers care

Social Search interviews reward freshness vs relevance trade-offs, privacy-aware ranking, and honest capacity math—not a generic Elasticsearch rectangle.

Key Highlights

•Unique angle for sec-01: Problem Statement
•Posts + people + hashtags with privacy-aware filters
•Kafka-backed indexing with monotonic post versions
•p95 200ms query / 30s index lag targets

1	public record SearchCursor(String searchAfterScore, String docId, long sessionTtlMs) {
2	public SearchCursor {
3	if (sessionTtlMs < 0) throw new IllegalArgumentException("ttl");
4	}
5	}

1	def blend_scores(post: float, people: float, tag: float, intent: str) -> float:
2	weights = {"posts": 1.0, "people": 0.9, "hashtags": 0.7}
3	base = post * weights["posts"]
4	if intent == "people":
5	base = max(base, people * 1.2)
6	return base + 0.15 * tag

1	export function normalizeQuery(raw: string): string {
2	return raw.normalize("NFKC").trim().slice(0, 512).toLowerCase();
3	}