Facebook database design?

I have always wondered how Facebook designed the friend <-> user relation.

I figure the user table is something like this:

user_email PK
user_id PK
password

I figure the table with user's data (sex, age etc connected via user email I would assume).

How does it connect all the friends to this user?

Something like this?

user_id
friend_id_1
friend_id_2
friend_id_3
friend_id_N

Probably not. Because the number of users is unknown and will expand.

Solution 1:

Keep a friend table that holds the UserID and then the UserID of the friend (we will call it FriendID). Both columns would be foreign keys back to the Users table.

Somewhat useful example:

Table Name: User
Columns:
    UserID PK
    EmailAddress
    Password
    Gender
    DOB
    Location

TableName: Friends
Columns:
    UserID PK FK
    FriendID PK FK
    (This table features a composite primary key made up of the two foreign 
     keys, both pointing back to the user table. One ID will point to the
     logged in user, the other ID will point to the individual friend
     of that user)

Example Usage:

Table User
--------------
UserID EmailAddress Password Gender DOB      Location
------------------------------------------------------
1      [email protected]  bobbie   M      1/1/2009 New York City
2      [email protected]  jonathan M      2/2/2008 Los Angeles
3      [email protected]  joseph   M      1/2/2007 Pittsburgh

Table Friends
---------------
UserID FriendID
----------------
1      2
1      3
2      3

This will show that Bob is friends with both Jon and Joe and that Jon is also friends with Joe. In this example we will assume that friendship is always two ways, so you would not need a row in the table such as (2,1) or (3,2) because they are already represented in the other direction. For examples where friendship or other relations aren't explicitly two way, you would need to also have those rows to indicate the two-way relationship.

Solution 2:

TL;DR:

They use a stack architecture with cached graphs for everything above the MySQL bottom of their stack.

Long Answer:

I did some research on this myself because I was curious how they handle their huge amount of data and search it in a quick way. I've seen people complaining about custom made social network scripts becoming slow when the user base grows. After I did some benchmarking myself with just 10k users and 2.5 million friend connections - not even trying to bother about group permissions and likes and wall posts - it quickly turned out that this approach is flawed. So I've spent some time searching the web on how to do it better and came across this official Facebook article:

TAO: Facebook’s Distributed Data Store for the Social Graph
TAO: The power of the graph.

I really recommend you to watch the presentation of the first link above before continue reading. It's probably the best explanation of how FB works behind the scenes you can find.

The video and article tells you a few things:

They're using MySQL at the very bottom of their stack
Above the SQL DB there is the TAO layer which contains at least two levels of caching and is using graphs to describe the connections.
I could not find anything on what software / DB they actually use for their cached graphs

Let's take a look at this, friend connections are top left:

enter image description here

Well, this is a graph. :) It doesn't tell you how to build it in SQL, there are several ways to do it but this site has a good amount of different approaches. Attention: Consider that a relational DB is what it is: It's thought to store normalised data, not a graph structure. So it won't perform as good as a specialised graph database.

Also consider that you have to do more complex queries than just friends of friends, for example when you want to filter all locations around a given coordinate that you and your friends of friends like. A graph is the perfect solution here.

I can't tell you how to build it so that it will perform well but it clearly requires some trial and error and benchmarking.

Here is my disappointing test for just findings friends of friends:

DB Schema:

CREATE TABLE IF NOT EXISTS `friends` (
`id` int(11) NOT NULL,
  `user_id` int(11) NOT NULL,
  `friend_id` int(11) NOT NULL
) ENGINE=InnoDB AUTO_INCREMENT=2 DEFAULT CHARSET=utf8;

Friends of Friends Query:

(
        select friend_id
        from friends
        where user_id = 1
    ) union (
        select distinct ff.friend_id
        from
            friends f
            join friends ff on ff.user_id = f.friend_id
        where f.user_id = 1
    )

I really recommend you to create you some sample data with at least 10k user records and each of them having at least 250 friend connections and then run this query. On my machine (i7 4770k, SSD, 16gb RAM) the result was ~0.18 seconds for that query. Maybe it can be optimized, I'm not a DB genius (suggestions are welcome). However, if this scales linear you're already at 1.8 seconds for just 100k users, 18 seconds for 1 million users.

This might still sound OKish for ~100k users but consider that you just fetched friends of friends and didn't do any more complex query like "display me only posts from friends of friends + do the permission check if I'm allowed or NOT allowed to see some of them + do a sub query to check if I liked any of them". You want to let the DB do the check on if you liked a post already or not or you'll have to do in code. Also consider that this is not the only query you run and that your have more than active user at the same time on a more or less popular site.

I think my answer answers the question how Facebook designed their friends relationship very well but I'm sorry that I can't tell you how to implement it in a way it will work fast. Implementing a social network is easy but making sure it performs well is clearly not - IMHO.

I've started experimenting with OrientDB to do the graph-queries and mapping my edges to the underlying SQL DB. If I ever get it done I'll write an article about it.

How can I create a well performing social network site?

Update 2021-04-10: I'll probably never ever write the article ;) but here are a few bullet points how you could try to scale it:

Use different read and write repositories
Build specific read repositories based on faster non-relational DB systems made for that purpose, don't be afraid of denormalizing data. Write to a normalized DB but read from specialized views.
Use eventual consistence
Take a look at CQRS
For a social network graphs based read repositories might be also good idea.
Use Redis as a read repository in which you store whole serialized data sets

If you combine the points from the above list in a smart way you can build a very well performing system. The list is not a "todo" list, you'll still have to understand, think and adept it! https://microservices.io/ is a nice site that covers a few of the topics I mentioned before.

What I do is to store events that are generated by aggregates and use projects and handlers to write to different DBs as mentioned above. The cool thing about this is, I can re-build my data as needed at any time.

Solution 3:

Have a look at the following database schema, reverse engineered by Anatoly Lubarsky:

Facebook Schema

Solution 4:

My best bet is that they created a graph structure. The nodes are users and "friendships" are edges.

Keep one table of users, keep another table of edges. Then you can keep data about the edges, like "day they became friends" and "approved status," etc.