Tinder

Location-based dating app

Design a dating app that shows each person a stack of nearby profiles, records a like or pass on each, and creates a match the moment two people like each other.

Overview

A swipe-based dating app shows each person a stack of profile cards, one at a time. The person likes or passes on each card. When two people have liked each other, the app announces a match and opens a conversation.

The stack is a search problem: find people within the user's chosen distance who fit both sides' preferences (age, gender), whom the user has not seen before, and put the most promising first. The swipe is a write problem: billions of small writes a day, and a check on every like for the other side's like. Location ties them together: the search index is split by geography, and people move.

Questions to ask first

How many users and swipes? — 20 million daily users, 100 swipes each a day
How far can people search? — Up to 100 miles, set per user
Must a match be instant? — Yes: the second like returns the match in its response
Can a profile be shown twice? — No, once swiped it never comes back
How is the stack ranked? — By likely mutual interest; the model is out of scope, the pipeline is in
Moving to another city? — Yes; the user must show up there within seconds

What you will be able to do

Find nearby users with S2 cells and a covering of the search circle.
Split a search index by geography so the shards carry equal load.
Build a card stack: candidates, filters, the seen filter, ranking, caching.
Detect a mutual like exactly once, even when both likes arrive together.
Keep the index correct when users move between shards.

Key concepts

Geospatial index
S2 cells
Geosharding
Load score
Candidate generation
Seen filter (Bloom filter)
Mutual match
Atomic upsert
Write ordering

Functional requirements

Create a profile with photos, age, gender and the preferences you want to see
Get a stack of nearby profiles that fit both people's preferences
Like or pass on each profile; never see it again
Get a match, with a notification, when a like is mutual
Update location; appear to people in a new city

Non-functional requirements

A swipe returns in under 100 ms at p99, including the match check (a target we chose)
A card stack loads in under 300 ms at p99 (a target we chose)
Two mutual likes produce exactly one match, even at the same instant
A user who moves appears in the new area within about 10 seconds (a target we chose)
A profile already swiped is never shown again

Interview questions

How do you find people within 20 miles quickly?
How would you shard the search index?
Both people swipe right at the same moment. What happens?
How do you make sure nobody is shown twice?
How is the stack ranked? Is it Elo?
A user flies to another city. How do they show up there?
What is the heaviest load, and how do you handle it?