Scaling to one million RPS

This Forbes Article and its Original Post describe how a guy managed to hit 1Million RPS for a while with one load balancer and 10$. This post also contains a gist to reproduce the experiments, which is pretty cool. This Facebook blog shows how they scaled up to 500 million users in 2010. In this article, which is also from 2010, Facebook gives an insight about their statistics and scaling strategies. Stories like these excite me to think about future scaling of Communify. With 500 Million users, Facebook had 13M requests per second. Now, with 4 billion users, they probably handle ±100M requests per second. I wonder how proper scaling could end up there.

My Geo Scaling Plan: scaling without bottlenecks

Since we have global communities now, it's pretty much impossible to get one user's data from one server. Therefore, the load balancer may need to have an overview of which communities are housed where. Then, the load balancer can send the query to the right server, based on the request.

If this wasn't the case, it would be easy, like described in my previous article about geo-scaling. But now that we have global communities too, we have a few tricks to do. This is what I came up with:

  • PULL automatically pull my GitHub repo everywhere whenever a new commit is pushed. pull, restart pm2, and automatically merge db changes to the db.

  • SPLIT automatically split a server into 2 servers whenever load is too high. Shard on community. The server is split in two, so there should be drawn a line on a map (cluster by location) so that the total members are split in two, based on location. There is an edge case in which an user is member of a community in both clusters. In this case, put the user on both servers. This is no big deal. The new server should then let the load balancer know it exists, and expose its communities.

  • MERGE automatically merge a server with another server whenever load is too low. Because we merge servers, we can't use incremental ID's, but need a UUID or GUID to keep table-rows unique. Since the sharding/splitting was done on communities, we can just merge all rows in all tables together. But because users may be in both clusters, there can be double users. If this happens, pick the user that's last updated, because that's where the user was active last. The deleted server should let the load balancer know it got deleted, so it will be deleted from the global server list.

  • BALANCE put a load balancer in front of all of it that directs every request to the right server, based on communityid. it knows which server has which communities because every server exposes their communities by communities(){id} and we know every server, so, every minute or so, we can look up all community-id's from all servers. This is pretty light. However, this can also be done the other way around. When a community gets created or removed, or when a server splits up or merges, the load balancer gets notified with the new composition. This would be instant and way cheaper.

Double Models

Models that can be sharded based on community, and thus just need to live on one server, and have just one copy: Posts, Subs, Roles, Communities, Channels.

Models with some problems: Users, CommunitySubs, Locations

A user can be subbed to two communities that live on different servers at the same time. There are a few possibilities to deal with this:

  1. Copy/paste When a user updates the current community to a community on a different server, copy that user, together with all of its CommunitySubs and Locations to the new server. This will, then, be the single server that user gets its data from. CommunitySubs get notification increments by mutation calls (that increment) from other servers where the user still exists. When a user changes community, all servers should be notified, so that if a server knows about a user, it also knows in which community that user is... This can get heavy, but it doesn't happen that often. A side effect of this strategy is that users, communitysubs and locations can get outdated on servers on which the user isn't active. However, all servers that need to know, know on which server a user currently lives, all servers can get updates about that user to stay updated. For example, every hour, or every day. I don't know how much will be useful. In principle, a user doesn't change much for another community if it's not active in that community.

  2. Global, seperate database for users, community and locations This can be nice because its a single source of truth with is always up to date. However, the drawback is having multiple servers the app has to connect to, and there is one global server, which is bad for availability (risk) and can't scale infinitely.

I think this is where I have to choose from, and I think option one is the best. I still have to discuss this with an expert. I'm quite impressed with this idea, because my app can scale infinitely big without bottlenecks based on a few assumptions, which my design of the app can guarantee:

  1. It doesn't get bigger than one loadbalancer can handle (around 1M rps)
  2. A single community never has to be sharded

This whole architecture is, I think, very interesting and would also work for my Chat-BaaS idea.

From 1M to 100M RPS.

100 M RPS, 4.000 M users, and ±40 M communities, ±20.000 servers.. That's the dream!

On a single server that's balanced well, I don't think any problems will arise. The only problem is that, when a user changes community, all servers that know this user should know this, so the server has to send 20.000 requests! Right? Well it does. Unless it knows on which server all other communities the user is subbed to are hosted. And the load balancers know this, right? So let's ask a load-balancer, and then just let the servers know that care! Great! Problem solved.

The other thing is, we have to balance the traffic with load balancers. One load balancer doesn't cut it anymore. 100 times as many RPS should mean 100 times as many load balancers, but because we have some extra work for the load balancers (telling servers which communities are hosted where), I think 1000 load balancers would be better, just to be sure. With 40M communities, is it still doable to let all load balancers know which communities are hosted where?

  • If yes, the problem gets easy. Just have one 'Master Load Balancer' that assigns any new visitor to one of the 1000 Load Balancers, and keeps it there. From there, the visitor knows where to go, because every Load balancer knows everything. Searching from ±40M rows of communities could be doable, but the bottleneck is probably that, on every user changing community (that is on a different server), every Load Balancer gets notified. The question is: How many times per second does this happen with 4 Billion users? If this is more often than ±5.000, the bottleneck is too small, and it won't fit. We have to find another solution.

  • If no, it gets complicated. Splitting the communities over all load balancers may be an option, but it's a complicated mess that I would have to think through. Let's save that for another time!

Scaling to one million RPS
Share this

Subscribe to Karsens