Scaling database with Django and HAProxy

MySQL – Primary data store

At HackerEarth, we use MySQL database as the primary data store. We have experimented with a few NoSQL databases on the way, but the results have been largely unsatisfactory. The distributed databases like MongoDB or CouchDB aren't very scalable or stable. Right now, our status monitoring services use RethinkDB for storing the data in JSON format. That's all about using NoSQL database for now.

With the growing amount of data and number of requests every second, it turns out that the database becomes a major bottleneck to scale the application dynamically. At this point if you are thinking that there are mythical (cloud) providers who can handle the growing need of your application, you couldn't be more wrong. To make matters worse, you can't spin a new database whenever you want to just like you do with your frontend servers. Achieving horizontal scalability at all levels requires massive re-architecture of the system while being completely transparent to the end user. This is what a part of our team has focused on in the last few months, resulting in high uptime and availability.

It was becoming difficult for the master (and only) MySQL database to handle the heavy load. We thought we will delay any scalability at this level till the single database could handle the load. We would work on other high priority tasks instead. But that wasn't to be, and we experienced some down time. After that we did a rearchitecture of our application, sharded the database, wrote database routers and wrappers on top of django ORM, put HAProxy load balancer infront of the MySQL databases, and refactored our codebase to optimize it significantly.

The image below shows a part of the architecture we have at HackerEarth. Many other components have been omitted for simplicity.

Database slaves and router

The idea was to create read replicas and route the write queries to the master database and read queries to slave (read replica) databases. But that was not simple either. We couldn't and wouldn't want to route all the read queries to slaves. There were some read queries which couldn't afford stale data, which comes as a part of database replication. Though stale data might be the order of just a few seconds, these small number of read queries couldn't even afford that.

The first database router was simple:

All the write and read queries go to the master database, which you might think is weird here. Instead, we wrote getfromslave(), filterfromslave(), getobjector404fromslave(), getlistor404fromslave(), etc. as part of django ORM in our custom managers to read from slave. So whenever we know we can read from slaves, we call one of these functions. This was a sacrifice made for those small number of read queries which couldn't afford the stale data.

Custom database manager to fetch data from slave:

HAProxy for load balancing

Now there could me many slaves at a time. One option was to update the database configuration in settings whenever we added/removed a slave. But that was very cumbersome and inefficient. A better way was to put a HAProxy load balancer in front of all the databases and let it detect which one is up or down and route the read queries according to that. This would mean never editing the database configuration in our codebase — just what we wanted.

A snippet of /etc/haproxy/haproxy.cfg:

The configuration for the slave in settings now looked like this:

But there is a caveat here too. If you spin off a new server with the HAproxy configuration containing some endpoints which don't exist, HAproxy will throw an error and it won't start, making the slave useless. It turns out there is no easy solution to this, and haproxy.cfg should contain existing server endpoints while initializing. The solution then was to let the webserver update its HAproxy configuration from a central location whenever it starts. We wrote a simple script in fabric to do this. Besides, the webserver already used to update its binary when the spin off is from an old image.

Database sharding

Next, we sharded the database. We created another database — analytics. It stores all the computed data, and it forms a major part of read queries. All the queries to the analytics database are routed using the following router:

To enable the two routers, we need to add them in our global settings:

Here the order of routers is important. All the queries for analytics are routed to the analytics database and all the other queries are routed to the master database or their slaves according the nature of queries. For now, we have not put slaves for analytics database but as the usage grows that will be fairly straightforward to do.

At the end, we had an architecture where we could spin off new read replicas and route the queries fairly simply and had a high performance load-balancer in front of the databases. All this has resulted in a much higher uptime and stability in our application, and we could focus more on what we love to do — building products for programmers. We already had an automated deployment system in place, which made the experimentation easier and enabled us to test everything thoroughly. The refactoring and optimization that we did in codebase and architecture also reduced the server count by more than two times. This has been a huge win for us, and we are now focusing on rolling out exciting products in the next few weeks. Stay tuned!

I would love to know how others have solved similar problems. Do give suggestions and point out potential quirks.

P.S. You might be interested in The HackerEarth Data Challenge that we are running.

Follow me @vivekprakash. Write to me at vivek@hackerearth.com.

This post was originally written for the HackerEarth Engineering blog by Vivek Prakash.

 

About the Author

Guest Author
Our guest articles are a collection of the best contributions made by members of the developer community on our blog. Discover articles on a wide range of topics, shared by top programmers across the world.
Share