ponyfoo.com

Low-Budget Hosting Tips

Fix
A relevant ad will be displayed here soon. These ads help pay for my hosting.
Please consider disabling your ad blocker on Pony Foo. These ads help pay for my hosting.
You can support Pony Foo directly through Patreon or via PayPal.

These are short-form “thoughts”, in addition to the usual longer-form articles in the blog. The goal is to publish one of these every weekday. I’d love to know what you think. You may send your questions to thoughts@ponyfoo.com. I’ll try to answer them over email and I may publish them here, with your approval. I also write thoughts about the current state of front-end development, and opinions on other people’s articles. You can use the form to the right (near the bottom in mobile) to subscribe via email.

I’ve been hosting Pony Foo on an AWS t1.micro instance for as long as I can remember. When it comes to AWS, “low-budget” is an oxymoron – as the cheapest type of instance on demand is priced somewhere near $30/mo. While the cheapest instance is pricey, it provides a far better experience than say, the lowest tiers of hosting on Heroku, which I’ve always found clumsy and quirky to use.

What I did like about Heroku back in the day I used it intensively was their approach to app development and how that translated into immutable infrastructure and deployments. I did learn tons and tons by following 12factor, and recently winded up writing up my own immutable deployment system, using Packer, the AWS CLI, and plain old bash.

All things considered though, I’ve learned far more from using IaaS-style hosting platforms like AWS than I did from Heroku. Using AWS forced me to learn more bash commands and UNIX_-y_ things than I ever had to worry about with Heroku. Learning how to come up with immutable deployments was also invaluable as it also gave me insight into how PaaS platforms work under the hood. Being able to tweak the instance exactly to your liking is invaluable too, and most PaaS really limit what you can do with their instances.

Here, I’ll share a few tips and tricks I’ve found useful when running applications using entirely custom scripts.

Add Some Swap Space!

In AWS, t1.micro instances have around 8GB of disk space while they only have around 800M of RAM. That’s very low RAM. I’ve incorporated the snippet below into my deployments to add 2GB of swap space, so that when the instance runs out of RAM, it can fall back to using the disk. Albeit slower, it’s often enough to get you there – considering you’re using a t1.micro instance or equivalent on another platform, you’re probably hosting a blog such as this one, a personal website, or a similarly low-budget project.

The script below will add 2GB of swap space on disk, mount it for immediate use, and make it so that it works across restarts.

echo "creating swap space"
sudo dd if=/dev/zero of=/media/swapfile.img bs=1024 count=2M
sudo mkswap /media/swapfile.img
sudo chmod 0600 /media/swapfile.img
echo "/media/swapfile.img swap swap sw 0 0" | sudo tee -a /etc/fstab
sudo swapon /media/swapfile.img

I’ve struggled with this one for quite a bit, because Pony Foo didn’t originally consume as much memory, but after throwing in nginx, io.js, a cluster of Node instances, and a cron job that runs every 6 minutes, memory ramped up. It took me a while to figure out, and the biggest issue was that AWS CloudWatch (their monitoring system) doesn’t track memory consumption by default, so the problem didn’t immediately stand out. The node processes would just hang, no error message or anything, and stop serving responses. Meanwhile CPU would ramp up to 100% utilization, rendering the instance useless for a couple of minutes.

CloudWatch displaying CPU utilization spikes
CloudWatch displaying CPU utilization spikes

Luckily, adding those 2GB of swap memory promptly fixed the issue!

Keep your Cluster in Your Cores

A similar issue was rooted in the cron job. After fixing the memory issues, I added in a very simple cron job written in Node that would figure out whether there’s any articles scheduled for publication and publishes them if their publication date was in the past.

As soon as I deployed, I noticed that the CPU would spike every six minutes. Suspiciously enough, that’s how often I had set up the cron job to run. After some thinking, I realized what was going on. As I mentioned earlier, I was already running a cluster on the instance, with os.cpus().length workers or 2, whichever was largest. Throwing in an extra Node instance crippled the CPU. Yes, that surprised me a bit too, the job runs for like 6 seconds most of the time. My work around was to place the scheduler in an API endpoint and use curl against the cluster in the cron job.

This had a couple of benefits.

  • CPU usage doesn’t spike to 100% anymore
  • Job runs faster because it doesn’t have to parse and compile any JavaScript, nor open any database connections

Nowadays CPU utilization hardly ever goes over 30%. Considering this is a t1.micro instance, that’s low utilization. High fives all around!

CPU utilization today for ponyfoo.com
CPU utilization today for ponyfoo.com

It does have the drawback that the job could be executed by an unwanted party, but that could be easily mitigated by doing some awkward dance where we verify that the requester already had prior access to the server file system anyways. That being said, this job in particular happens to be idempotent – running it over and over again won’t change the outcome – and thus, not a big deal that people can help me check if an article slated for publication can already be published.

Free-tierize Your One-offs

If you have one-offs that need to be executed every once in a while, such as the cron job I’ve just described, maybe you could spin up free-tier instances on a PaaS provider to run those jobs. It’s a little extra work, but it lightens the load on your application servers and spares the CPU!

Monitor All the Things

When you’re running a website on a budget, monitoring plays an important role. Your site might go out of memory, grind to a halt and no longer be able to serve any responses, crash, time out on the database connection, and many other risk factors are always lurking around.

I have a love/hate relationship with Uptime Robot. On the one hand, it’s a great service that pings my sites every few minutes and emails me when they go down (and back up). On the other hand, getting those ponyfoo.com is down” emails sucks. Just look at those slick graphs, though!

Uptime Robot Dashboard for ponyfoo.com
Uptime Robot Dashboard for ponyfoo.com

Of course, you shouldn’t rely solely on Uptime Robot. Add logging to your application. Use winston to log messages to your database whenever an error occurs during a request, a worker in your cluster crashes, and whenever similarly scary events occur.

If you’re on AWS, you might want to hook into the CloudWatch API, or just head over to your dashboard and see if anything seems amiss.

CloudWatch Metrics in the AWS Dashboard
CloudWatch Metrics in the AWS Dashboard

If you’re not on AWS, you could always look to [New Relic] and similar “embedded” APM (Application Performance Management) solutions. These usually take their own toll on performance – much like “anti-virus” software (which I hate) – but they do have a noble purpose in thoroughly monitoring and providing you with real-time statistics and insight into the current state and load of your application.

Hope that helps, I probably forgot to mention some obvious advice here, but this is mostly just off the top of my head!

Have any questions or thoughts you’d like me to write about? Send an email to thoughts@ponyfoo.com. Remember to subscribe if you got this far!

Liked the article? Subscribe below to get an email when new articles come out! Also, follow @ponyfoo on Twitter and @ponyfoo on Facebook.
One-click unsubscribe, anytime. Learn more.

Comments (7)

Roman Boiko wrote

t2.micro would probably cost about twice less, and even cheaper if it was prepaid: about $150 for 3 years or $75 for 1.

Also, using Docker with something like alpine linux, many apps can be deployed as extremely small (~6MB for statically compiled services like nginx or between 20 and 200MB for nodejs) images that consume very little resources, and thus it is often possible to host several of them on the same instance.

Nicolas Bevacqua wrote

Wow. I did not know this. I’ll look into using t2.micro. Many thanks! You might have saved me $30 a month, haha!

Guilhem Soulas wrote

It’s certainly not worth the switch but DigitalOcean is a great IaaS for small projects like a blog or a personal website and it costs $5/month. It provides you with a Linux instance that you can SSH and that’s all, nothing else!

Kyle Hayes wrote

+1 for getting a hosted Linux server for learning more. Awhile back, I wanted to learn the ins and outs to hosting a website on Linux as well as learning how to do more from the command-line. After grabbing a slice from SliceHost, I learned more in that period of time because I was forced to configure my own iptables, install packages from apt-get, and learning how to configure everything together. These days, I work at a shop that primarily runs RedHat and OS X and the time I used to get up to speed on *nix certainly helped.

Martin Pitt wrote

I followed your strategy in kind of a reverse. Originally started with Rackspace and it was awesome being able to learn all the unix things but having to take so much control and freedom comes at a big price: you are now responsible for maintenance. You now need to constantly supervise your system at low level. So when I switched over to heroku I felt a huge burden lifted off my shoulders, I could concentrate on deploying web apps and nodejs side of things without worrying too much about the details of having to setup, maintain and configure servers. I still have a rackspace account though because I find that hosting static websites via rackspace CDN costs peanuts and can make for a crazy fast website.