Design Facebook Post Search
DifficultyMedium | HelloInterview: problem breakdown
Problem Statement
Design a search system that lets users search across their own posts and their friends' public/shared posts by keyword.
๐Real-world: The hidden hard part is that search and access control collide. A naive inverted index returns "all posts containing the word," but here the results must be filtered by who's allowed to see them โ your friends' friends-only posts, privacy settings, blocks โ and doing that filtering after ranking breaks pagination and leaks existence. Facebook's actual system (Unicorn, their graph-aware search engine) had to make the index itself privacy-aware. This is the core security lesson for the audience: search is a notorious authorization-bypass surface โ if you forget to apply the viewer's permissions to every result, you've built a data-leak engine, and "search returned something I shouldn't see" is a classic real bug class (IDOR-via-search). Technically it's an inverted index (Elasticsearch) + async indexer (Kafka) re-ranked by recency and social signals, but the headline move is to bake per-viewer access control into retrieval, not bolt it on after.
Requirements
Functional
- Full-text search across posts
- Results filtered by: my posts, friends' posts, public posts
- Ranked by: relevance + recency
- Support for typo tolerance
- Privacy-aware: only show posts the searcher has permission to see
Non-Functional
- Billions of posts indexed
- Search latency: < 500ms
- Index freshness: new posts searchable within 1 minute
- Privacy enforcement: strict โ never return unauthorized posts
Core Design
Search Index with Privacy
The hard part isn't the search โ it's enforcing privacy at search time without re-querying the DB for every result.
Approach 1: Pre-filter index โ each document in the search index includes a visibility list (which user IDs can see it). Query filters by visibility.
- Problem: visibility list can be huge (public posts visible to all). Storing a list of 2B users in an index document doesn't work.
Approach 2: Search then post-filter โ search returns candidates, then DB lookup enforces privacy.
- Problem: if top 20 results are all private, you get 0 results after filtering. Need to fetch more candidates.
Approach 3: Visibility categories (recommended):
Post visibility types:
- SELF: only author sees it
- FRIENDS: author + friends list
- PUBLIC: everyone
Index documents with visibility=PUBLIC are searchable by all.
Index documents with visibility=FRIENDS: index separately; only search this sub-index when the searcher is a friend of the author.At search time: query the public index + the friends-only indexes of all the searcher's friends (or a subset if too many friends).
Write Path (Index Update)
New post created โ Post Service โ Kafka โ Search Indexer
โ
Elasticsearch (index document)
Document contains: post_id, author_id, content, created_at, visibilityRead Path
Search query โ Search Service
โ
1. Build permissioned query:
visibility=PUBLIC
OR (visibility=FRIENDS AND author_id IN [user's friend IDs])
2. Execute against Elasticsearch
3. Rank: BM25 relevance ร recency boost
4. Return with post metadata (from Post Service cache)Friend IDs at Search Time
"User's friend IDs" โ up to 5000 friends. Passing 5000 IDs to Elasticsearch filter is doable. For users with many friends, pre-compute a "friends set" in Redis per user.
Architecture
Post Write โ Kafka โ Search Indexer โ Elasticsearch
โ
Search Query โ Search Service (add friend IDs from Social Graph)
โ
Elasticsearch (permissioned query)
โ
Hydrate post details โ ResponseSecurity Considerations
| Threat | Mitigation |
|---|---|
| Privacy bypass | Visibility category in index; friends-only posts only returned when friendship verified |
| Content injection via search | Sanitize search query before Elasticsearch query construction (prevent injection) |
| PII in post content | Content moderation; GDPR right to erasure must propagate to search index |
| Enumeration attack (search to find private posts) | Consistent 0-result response for unauthorized posts; no timing difference |
๐ฐReal incident โ Facebook phone-number scraping (2019, leaked 2021). Attackers abused Facebook's contact-import feature โ which matched uploaded phone numbers to accounts โ by submitting huge ranges of numbers and recording which profiles came back. Facebook closed the gap in 2019, but the scraped data on about 533 million users was posted free online in 2021. Any search or lookup that maps a private attribute to an identity is an enumeration oracle unless it's tightly rate-limited.
๐ฏOn the job. Check that search results are filtered by the viewer's permissions inside the search index query, not afterwards in application code. Post-filtering leaks through counts ("about 40 results" when you can see 3), through ranking, and through timing โ and through any code path that forgets to apply the filter.
Interview Tips
spend the most time on visibility categories and how you enforce them at index time vs query time.
standard choice; mention BM25 scoring and fuzzy matching for typo tolerance.
near-real-time via Kafka is better than polling DB.