Let's design the high-level architecture of Twitter. I mean, how hard could it be? By the way, this video is taken from my ongoing course, System Design Interview, which will be complete by the end of this month. You can check it out on neco.io. Before we get started, I do want

to mention Twitter has been quite a popular topic recently, especially the underlying infrastructure and design. But keep in mind that in a real interview, your design does not have to exactly match the product. That's not what it's about at all. It's about discussing the

trade-offs and kind of demonstrating your knowledge of being able to weigh the pros and cons of an approach. And of course, there's many similar products to Twitter. There really isn't any one correct approach, so we don't have to actually replicate the real Twitter design. Unless, of course, you're

interviewing at Twitter. In that case, you might have to, because they recently fired everyone, so they need people to know how it works. So let's start with the background. We know that Twitter is a social network first and foremost, where some people can follow other people, and

that relationship can be mutual. This person can also follow the other person, but some people might end up with more followers than others, right? So assume that one person is really popular and everybody wants to read all of their tweets, but you know, most people on

Twitter probably aren't actually tweeting very often. Most people don't actually have many followers, including myself. I'm mentioning this because it kind of hints that this is going to be a very read-heavy system. And of course, on Twitter, the whole point is that people

can create tweets so on a particular tweet you have a person like this is their profile picture and their username and then the actual content of the tweet it can have some text it can have some images and it can have a video there's a lot of things you can do to interact

with a tweet of course you can like the Tweet you can do a retweet you can follow the person who actually made the Tweet or unfollow them recently you can even edit tweets now but this is just to give you a general idea of what kind of

functionality you might want to clarify with your interviewer so now actually digging into the functional requirements Twitter is very very large of course we can't design every little piece of functionality in a 45 minute interview so what exactly do we want to spend most

of our time on and what parts can we just kind of hand wave and dismiss and just briefly discuss let's say the first feature the priority feature is that we want to be able to follow other users so users can follow each other now there's no point of following other people if

you can't actually create tweets so that's also going to be just as important and then third is actually viewing a news feed now at a basic level these two features are pretty simple but viewing a feed can definitely be more complicated especially when we get into

how we want to rank that feed what kind of algorithm are we going to use probably there's machine learning going on in there and in many cases you end up seeing Tweets in your feed by people that you're not even following but we're gonna assume that that's not the case

for reviewing a feed we just want to see tweets of people that we actually follow I think that would be something worth clarifying with your interviewer that's something that can kind of scope down this interview because you might assume that we're doing something really complicated but your interviewer is

actually looking for something more simple that's a trap that you don't want to fall into now what exactly is going to go in the Tweet itself we know that Twitter actually has a limit on tweet size I think it's 140 characters let's assume that your interviewer gives you

that number but at the same time with social networks of course we end up with images and videos and let's say that yes we are going to include these in our design so now let's transition into the non-functional requirements which is not going to be completely separate from

these actual features we're implementing the first thing you probably want to know is how many users we're talking about here let's say the number is 500 million total users but in terms of daily active users we have about 200 million of them that are daily active I

think that's pretty close to the real number but remember the real number isn't so important I think the main observation here is that almost half, half of the users are active so when we actually create feeds for users we'll be doing it for most users most people are

going to be logging on most people are going to be viewing their feed well not most but nearly half which is a pretty large percentage about 40 percent Now While most people will be viewing their feed they probably won't be creating tweets but again this is something we

have to clarify let's say of those 200 million daily active users each of them will read about a hundred tweets per day so 200 million times a hundred that's going to be 20 billion tweet reads per day now what is the size of each tweet

if we have 140 characters that's about 140 bytes but let's assume that there's additional information with a tweet we have the username of that tweet and possibly there's a lot more metadata to be safe or just a basic tweet that just includes text we can assume that for

each tweet we have to do a kilobyte of reading from our storage now we also know that some tweets can contain images and videos so on average this is going to be higher how much higher we could spend a lot of time digging into the

math you could ask your interviewer a bunch of questions but most likely this is not what they want you to spend time on let's just average this up to a megabyte because maybe videos on average are 10 megabytes if they're a bit longer

which I don't know what the limit is for a video length on Twitter but it could be reasonably high but we also know that few tweets are going to actually have this so we average it down to this because most tweets are going to be about a kilobyte let's say so let's say

a megabyte per each tweet so how much data are we going to be reading if it's 20 billion tweets a megabyte for each tweet that is quite a lot of data so if each tweet was just one byte 20 billion that's going to be 20 Gigabytes but now

we're actually multiplying this by a megabyte which is a million so multiply this by a thousand we get 20 terabytes but we have to multiply it by a thousand again because that's you know what a million is and then we get to 20

petabytes so overall we're going to be reading 20 petabytes of data per day now it's no surprise to us that this is going to be a read heavy system this is kind of hinting to us what type of storage solution should we use we

probably don't need to be strongly consistent eventual consistency is enough and that brings us to how much are we going to be writing per day how many tweets are we going to be creating per day well we have 200 million daily active users so let's say a reasonable

number is 50 million tweets created per day most people aren't going to be creating tweets but maybe some people create 10 tweets per day so this is a decent number but you're not going to be guessing this your interviewer should be giving you something reasonable now we

could go through the rest of the math with this number and I can show you that we're going to be writing much less than 20 petabytes of data per day especially if we don't include the images and videos which we're probably not going to

be directly storing in a database but the main thing here to realize is that yes we're going to be writing much less than we're reading so that's how we want to optimize our design and let's say that the average user follows about a hundred people so a hundred follows

per person, but of course there can be power users who have a thousand or ten thousand people that they follow. But the more important consideration here is for a user, how many followers can they have? Someone like Kim Kardashian, I don't know if she's the most popular on Twitter, but

I think it's at least over a hundred million followers that they have. So this is the more important consideration, people who have a massive amount of followers. So the question is going to be, for all the people that follow Kim Kardashian, how are they gonna get the

tweets? This is kind of hinting to us that wherever we're storing her tweets, it's gonna get overloaded pretty quickly. So now that we kind of know what we're looking for, let's get into the high-level design. We know of course that everything is going to start with our

client, whether that's a computer or a mobile device, it doesn't really matter for us, we're focusing on the back end, which is agnostic to the front end. We know the first thing our user is going to be hitting is the application servers to perform actions like creating a tweet

or getting their news feed or following someone. Now because of the scale that we're dealing with, we're probably going to be bottlenecked by getting the news feed. That's what's going to be happening most frequently, and if we want to scale this up, assuming that these application

servers are stateless, it should be easy to scale them up, and of course we will have the load balancer in between this. That's pretty hand-wavy. I mean, that's something you can just memorize and say, well if you want to scale horizontally

scale this and put a load balancer in there. It's pretty trivial, so I'm not going to spend a lot of time on that now. Of course, our application server is going to be reading from some storage. Let's say we do have a database, and

let's say that it is a relational database. And you might be thinking, if we're going to be doing read-heavy, why use a relational database? Why not just have a NoSQL database? Well, it depends on what type of data we're going to be storing. Do we need joins in this case?

And we could, because we do have a very relational model when it comes to following. That's a pretty clear relationship between followers and followees, so that's a reason to go with a relational database. Now, in theory, you it would be easier to scale a NoSQL

database, but we can implement sharding with a relational database, so that does give us some flexibility. Though, after finishing our high-level design, we might want to revise this because we could just store tweets and user information in a NoSQL database, and then have the

graph DB, which would be very easy to find that follower relationship, because a graph DB is essentially like an adjacency list graph where every person is like a node in a graph. And to find all the people that they follow, you just have to look at every outgoing edge, and

to find all the followers of a person, you just have to look at every incoming edge. So, depending on your expertise and your background, and of course, what your interviewer is looking for and what they might be familiar with, you can kind of have some discussion about these

differences now with the massive amount of reads that we're going to be doing, we basically have to have a caching layer in between. So as we're reading tweets, we will be hitting our cache before we hit our database. But also remember that we are going to be storing media, so we need

a separate storage solution for that media related data. National databases aren't the best for that, so we'll have some type of object storage for that, something like Google Cloud Storage or Amazon S3. So when we actually read a tweet, we'll be getting the information about that tweet, like the Tweet ID, who's

the creator of that tweet, what time was it created, whether it included an image or not, what was the image that it included, the profile picture of the person who made it. The application server can then fetch the image, like the profile picture or the video that showed

up in that tweet, and it can do that separately. But at the same time, because these assets are static in nature, it may be better to actually distribute them over a CDN Network, so then actually our application server does not have to interact with the object storage. The

application server will respond to the user with all the information that they need, including the URL of that image or video that they need, and then the client, whether they're using a mobile device or a desktop browser, will make a separate request which will actually hit our CDN

Network which is tied to our object storage. What type of algorithm would we use in this case? Well, even though we're looking at the high level right now, we probably want to use a pull-based CDN. We don't want to necessarily push every image or video to the CDN immediately

also remember the benefit of a CDN is that it's geographically located close to the user we know that people in India might be looking at different types of tweets and images and videos than people in the United States so it doesn't make sense to put every single new tweet push

it directly to the CDN Network and with a pull-based CDN we kind of guarantee that the stuff that's loaded onto our CDN is the relevant stuff that people want to see anyway the popular things so now let's spend most of our time actually digging into the details which

some people like to start out with the interface that we'll be using we will have a couple so remember we have a create tweet there could be a lot of metadata sent with that request of course the user ID of the person creating the Tweet but mainly the user

is actually responsible for sending the content of the tweets so one is the actual text and then second is going to be the actual media of course every tweet has a created timestamp but we assume that that will be handled server side and every tweet has to be

identified but we'll assume that the Tweet ID is also created server side and of course the user ID of the person that's actually creating it well we'll assume that in the header of the HTTP request that there's some authorization token for us to know that the correct

person is making the Tweet but we could also have the user ID passed into this request or the username of the person and next of course we have getting the actual feed and that really doesn't need any information at all that's a very basic read request we don't need to send

additional data that we're going to be actually storing. Get feed should just pass in the user ID so that we know which person's feed are we getting. But at the same time, I should not be able to pass in your user ID even if I know it.

to get your user feed. There's nothing that can go super wrong with that, but it shouldn't be allowed. And that would be handled by the HTTP header. We know that there's actual authentication going on in this system, it's just that we're

not focusing on that. We're focusing on the actual Twitter design. Authentication happens with pretty much every application, and we also have the follow interaction. So a user can follow another person, they'll pass in their user ID and maybe the username of the person that

they're trying to follow. So these are the three main interactions. Now, how are we actually going to be storing this data? In particular, we're going to be storing two things: the actual tweets. So assuming we have a relational database, we'll have a table of tweets, and we'll

also have a table of follows. So the follow relationship is pretty simple: you have the followee, the person who's following the other person. That's going to be a string, and we'll have another, the actual follower. I think I misspoke when I was describing followee. So the

follower is the person that's following, the following, the followee is the one that's being followed. Now, assuming that this is a table, and remember for a user, we want to get all the tweets of people that this person follows. So they're the follower, and they want all the tweets of

their followers. How would we index this table? We'd probably want to index based on the follower because then all records in that table will be grouped together based on the follower, so all the people that this guy follows will be grouped together in the table. It'll just be a

range query, so assuming this is our table, all records for this person will be grouped together in a particular range over here. I think it's worth mentioning that we would favor indexing based on the follower if we have a read-heavy system. Now for the Tweet itself, we

kind of briefly talked about it. Of course, we're going to have the Tweet ID, we're going to have the timestamp, we're going to have the user ID of the person who created it, and we're going to have the content of the Tweet, whether it's the

text, which technically could be empty if we have some media attached to it, but we actually won't be storing the media itself in the database. We'll have a reference to that media that references the object store. Now the bigger problem with our storage is we're going to be

storing a massive amount of data. If you went through the calculation that I mentioned earlier, you would have gotten to I think roughly 50 gigabytes of data are going to be written per day to our relational database if we're not including the media. So that's a lot of

data. In a month we'll have I think 1.5 terabytes. In the course of a year we'll have roughly 18 terabytes, so it is a large amount of data, but it's actually reasonable. And the good thing about Twitter is that tweets are relatively

small but the problem is we're going to be having so many reads hitting this database a person is going to be reading a hundred tweets per day and we have 200 million of them so the first approach and the obvious thing to do is to have

read-only replicas of this database if reads are the bottleneck it's not hard to just add additional database instances now the problem will be when a user actually creates a tweet if we have single leader replication all those writes are going to be hitting a single database and remember if we have 50

million writes per day you divide that by a hundred thousand which is roughly the amount of seconds in a day we get 500 writes per second well probably even more than that but at the same time there's gonna be moments where we have traffic this is the average but peak

could be much higher I mean peak could be even 10 times this amount depending on what's going on maybe Elon Musk does something crazy so ideally we should be able to scale our writes as well and the obvious way to do this is by using

sharding but the question is how are we going to be implementing this sharding before we get into that I forgot to mention if we do have read-only replicas it's okay if a single instance gets the writes and then asynchronously populates the read-only replicas with the data

because in the case that a user ends up hitting one of the replicas and gets some stale data it's okay if it takes five seconds after a tweet is created before a user actually gets that tweet or it might even take 20 seconds that wouldn't be ideal but it's not the

end of the world with something like Twitter now the question is how are we actually going to be sharding this what type of shard key are we going to be using I think the most obvious and easy way would be to do it based on user ID

because that's kind of the whole point of our design the way we scoped it out a user only cares about a subset of users they don't care about every user it doesn't make a lot of sense to do it based on tweet ID because what we want

is to break up our database into pieces and we want a particular user to ideally only have to hit one of these pieces with the logic of our system especially the sharding logic a user should know which people they follow and then we can route the request to the appropriate

shards that contain the tweets of the people that they follow and we'll know that because we can use our shard key to determine that but if we break up the tweets based on tweet ID we actually don't know which shard contains the tweets of the people that they follow so

we'd have to query all the shards that kind of defeats the purpose so we'll be choosing to do this based on user ID and since we don't actually have complex queries and joins that we're doing to actually retrieve the tweets how will we

get the people that they follow well if we shard that based on user ID as well all the people that this guy follows should be on a single shard now a potential problem is that we will actually have to order the tweets just fetching the tweets is not enough we

will actually have to query them and then order them based on the time that they were created and of course how many tweets are we looking for if we're just looking for a small number like 20, what if the user wants to scroll down in their feed and we want 20 more tweets, we

don't necessarily need to get all tweets immediately. We need to wait for the user to actually scroll, or maybe we can actually get all the tweets immediately, but before I even get into that, let's quickly look at how our current system is going to work when a user creates an

essentially, it will go through the server based on the user ID. We will find the appropriate shard and then store the tweet on that shard and then any images and media on object storage. As users view their feed, we may have to query multiple shards to find all the relevant

tweets and then order them and then send them back to the user. That could definitely be very slow. That's clearly the bottleneck here. Now we do have a caching layer, and in theory, the most popular tweets will be stored already on this caching layer, and if they're not

then we can have some type of LRU algorithm working here because we care about the most recent tweets. Most likely people aren't going to be viewing tweets from a year ago, even if they had like a billion views. After a while, people get tired of them. LRU might be better here

and in theory, caching should definitely help us lower the latency, but remember different people will have different tweets. We definitely can't guarantee that all 20 tweets that this person wants to see are already going to be cached. What if 19 of them are already

cached and then that last one tweet we still have to go and read disk to get that tweet and while you're scrolling maybe all your tweets are loaded but one tweet in the middle is taking a few extra seconds that's not a great user experience so the problem we're running

into is not really scalability it's latency and caching helps with that and sharding also helps with that that's kind of the point of sharding with read-only replicas we're able to handle scale but as we break up the data into smaller chunks then we can also lower

latency because we're going to be querying a smaller amount of data but we still may have to query multiple shards to get this latency even lower we can get pretty creative and we can actually generate the news feed of users asynchronously and we would do that for

every single user in theory because we know a large amount of the users 200 million out of 500 million are actually active it makes sense to generate the news feed for all of them even if 60 percent of them aren't going to actually view it it's not a ton of wasted work

and we could also tune it such that we only regenerate the news feed for people that are actually active within the last 30 days now at a very high level what we can do is have some kind of message queue or Pub/Sub system which will take every new tweet that is created will

also not just be written to the database but it will be sent from the app servers to the Pub/Sub queue and this queue will feed into a cluster something like a Spark cluster but the point is that these workers will in parallel process all the messages that we're getting

which will include every time there's a new tweet that is generated, there could be a lot, so we need to do this asynchronously. And the point of this is that these will basically feed into a new cache, and this cache will be responsible for actually storing the

user feed. So now when a user loads their home page and gets their list of 20 tweets, the application server will actually be hitting this feed cache. Maybe this cache is used for individual tweets that are maybe embedded on websites or sometimes you just open up

an individual tweet, not in the context of a user feed, and you want to maybe look at the replies of that tweet. But this is starting to make a little bit less sense, or even having this cache, so I'm just kind of giving some possible use cases, but this is the cache that

will actually have a feed for every single user. So if this is in memory, it may have a large amount of data. Now we have 200 million users, and if we want to have 100 tweets for every single user, it could be a very large amount, so we

probably want to shard this similar to how we did with our relational database, but the point is that this will definitely lower the latency because the feeds are not generated as a user actually requests to actually get the feed. We don't have to actually run a

complex query on our relational database, having to query multiple shards and then joining the results together based on the time that they were created and then ordering them like that. The feed will actually already be created. Now the complicated part is actually updating the tweet. We kind of went through the

flow of when a tweet is created, it will be added to the message queue, and then for that individual tweet, these workers will add that tweet to all the feeds of people that are following the author of that tweet. But now the problem is if somebody has a hundred followers, then

these workers will have to update a hundred feeds. But what about somebody like Kim Kardashian who has maybe a hundred million followers? Updating a hundred million feeds every single time somebody popular like that makes a tweet is very, very expensive. Maybe in that case, it's not the end of the world for

us to actually have to update the feed of that user at the time that they actually request it. Big, somebody could have a hundred million followers and us having to update 100 million feeds every time they make a tweet is pretty expensive. But not all 100 million of

those followers are even loading their feed every single day, so it'd probably be easier to do that work as it's needed. So when a user makes a request, they get their feed, but maybe in parallel to that, our app server could look for other tweets, probably a popular tweet by

Design Twitter - System Design Interview

Design Twitter - System Design Interview