|||

Video Transcript

X

Scaling ipify to 30 Billion Requests and Beyond on Heroku

The following is the story of how Randall Degges created a simple API to solve the common problem of external IP address lookup and how he scaled it from zero to over 10 thousand requests per second (30B/month!) using Node.js and Go on Heroku.

Several years ago I created a free web service, ipify. It is a highly scalable IP address lookup service. When you make a GET request against it, it returns your public-facing IP address. Try it out yourself!

Gif of ipify GET request returning public facing IP address

I created ipify because, at the time, I was building complex infrastructure management software and needed to dynamically discover the public IP address of some cloud instances without using any management APIs.

When I searched online for freely available reverse IP lookup services I didn’t find any suitable solutions:

  • There were websites I could attempt to scrape my IP from (but this is bad form, and would likely result in complaints from the host)
  • There were APIs for this that charged money (which seemed unreasonable for such a simple service)
  • There were APIs that allowed you to do a limited number of lookups per day (which scared me as I was managing a lot of instances at the time)
  • There were APIs that appeared to be what I wanted, but upon using them they’d error out, go down randomly, or just otherwise not be of high quality. When I inspected the dig records of a particular provider, I noticed that the entire service was running on a single server (with an A record) that was terminating requests directly: not the most scalable/highly available service in the world.
  • There was an API service that looked somewhat OK but was trying to raise money via donations to stay alive. Integrating with an API service on the brink of death didn’t make me feel terribly comfortable.

Because of all these reasons I figured I’d just throw a small service together myself and solve the problem for as many people as possible. After all, writing software to return a single string isn’t terribly hard. Why shouldn’t I do it?

I assumed that worst case I’d spend $30 bucks or so per month and treat it as a public service.

ipify v0

The first iteration of ipify was quite simple. I wrote an extremely tiny (< 50 LOC) API service in Node. Since the entire premise of the ipify service is returning a string, I figured this was a perfect use case for Node as a technology: handling a lot of simple requests with minimal CPU usage.

After I had the API service built in Node, I threw together a simple static site to power the front-end and deployed it into an S3 bucket on Amazon. I then configured a CloudFront origin (Amazon’s CDN service) to sit in front of the S3 bucket and cache the pages for ultra-fast load times.

Now, I’m not a designer by any stretch of the imagination… But, with a little bootstrap love things turned out half-decent.

Now, I needed to deploy it.

Enter Heroku

I’m a big fan of Heroku (I’ve even written a book about it). I’ve been using it for many years now and consider it to be one of the most effective services in the developer world.

I decided that if I wanted to run ipify in a scalable, highly available and cheap way then Heroku was the simplest and best option. So that’s what I went with.

I deployed ipify to Heroku in a minute or two, ran it on a single dyno, and did some limited testing. Things again seemed to be working well and I was feeling pretty happy with myself.

If you aren’t familiar with Heroku, let me explain how ipify’s infrastructure was working:

  • Heroku ran my ipify web service on a small dyno with 512M RAM and limited CPU
  • If my process crashed or had any critical issues, Heroku would automatically restart it for me
  • Heroku ran a load balancer that accepted all incoming requests for my app and forwarded them to my dyno to process the request

This is a nice setup because:

  • Everything is highly available: Heroku’s load balancers, my dyno, everything
  • It requires no maintenance, configuration management, or any sort of deployment code. It’s 100% automated.
  • It’s cheap: I paid ~$7/mo to run this single web server
  • It’s fast: Heroku runs on top of Amazon Web Services (AWS), so my infrastructure was running in one of the most popular cloud hosted destinations in the world: AWS.

At this point, I was feeling pretty good. This took less than 1 day’s worth of work to build, setup, test, and move into production. I then integrated ipify into my own infrastructure management code to solve the problem I had initially needed to resolve.

Things were going great for about a month before I started noticing some issues…

Popularity... Ugh

I never marketed ipify, but it ended up ranking really high for the “ip address api” search phrase on Google, and still does. I guess all those years working on copy editing and SEO paid off.

Within a month or so, ipify was ranking near the top of Google search results, bringing in thousands of new users. With the increased visibility of the service, I started seeing some issues. The Heroku load balancer was firing off warnings because my Node server wasn’t servicing incoming requests quickly enough. What ended up happening was:

  • Too many users would make API requests to ipify
  • My Node server would start responding to requests slowly so latency would rise
  • The Heroku load balancer would notice this and start buffering requests before sending them onto my Node server
  • Because my Node server wasn’t servicing requests quickly enough, the load balancer would return a 503 to the user and the request would die off

Not a pretty picture.

So what I did was simple: I added another Heroku dyno. This way, I’d have twice the capacity and things would run smoothly again. The downside is that scaling to two “production” dynos on Heroku increased my cost to $50. When you need to scale horizontally, you have to move from hobby dynos to standard 1X dynos starting at $25/dyno/mo.

I figured that by now the service had likely capped out in popularity and I was OK paying $50/mo, so I’d just do that and things would go back to normal.

But… Things didn’t quite work out that way.

Before even a week had passed, I was already getting alerts from Heroku telling me that I was having the same issue as before. When I looked at my stats, I could see that my traffic had doubled, and again it appeared like ipify was just simply getting too much usage. I added another dyno (brining my monthly cost up to ~$75/mo), but decided to investigate further. I’m a frugal guy – the idea of losing > $50/mo seemed unpleasant.

The Investigation

When I started investigating what was going on, the first thing I did was check to see how many requests per second (rps) ipify was actually getting. I was pretty surprised by the number: it was low.

ipify was only getting around 10 rps at the time. When I saw how small the number was, I figured something bad must be happening in my code. If I can’t service 10 rps across two small web servers, I must be doing something horribly wrong.

The first thing I noticed was that I was running a single Node process. This was an easy fix: I started using the Node cluster module, and bam, I was immediately running one process for each CPU core. This effectively doubled my throughput on my Heroku dynos.

But 20 rps still seemed like a tiny number so I did some more digging. Instead of doing load testing on Heroku, I did some load testing locally on my laptop.

My laptop was far stronger than a small 512M RAM Heroku dyno, so I figured I should see much better throughput.

I did some testing using the ab tool and was surprised to see that even on my laptop I was unable to surpass a threshold of 30 rps from my Node processes (I run Linux on my laptop and ab works effectively there). I then did some basic profiling and found that Node was spending a lot of time performing basic string manipulation operations (to extract the IP address from the X-Forwarded-For header and clean it up). No matter what I experimented with, I was unable to boost throughput up much above that limit.

At this point, the production ipify service was able to serve roughly 20 rps across two dynos. ipify was in total serving ~52 million requests/mo. Not impressive.

I decided to rewrite the service in Go (which I had started using a few months before) for performance purposes to see whether or not I could get more throughput out of an equivalent Go server.

ipify v1

Rewriting ipify in Go was a short (but fun) experiment.

It gave me the opportunity to mess around with many different Go routing stacks: Gorilla/mux, Martini, and httprouter. After benchmarking and playing around with all three routing tools, I ended up using httprouter as it performed significantly better than the other two more popular options.

On my laptop I was able to achieve ~2,500 rps from my Go server. A massive improvement. The memory footprint was also much lower and hovered around 5M.

With my new found love for Go, I immediately took action and deployed my new Go-based ipify service on Heroku. The results were fantastic. I was able to get about ~2k rps of throughput from a single dyno! This brought my hosting back down to $25/mo and my total throughput to ~5.2 billion requests/mo.

Several days later, while talking to a more experienced Go developer, I ended up rewriting some of my string handling functionality which netted me an extra ~1k RPS of throughput. At this point, I was able to sustain ~7.7 billion requests/mo per dyno (give or take a bit).

I was thrilled to say the least.

More Popularity

Although I was able to cut back my hosting cost to a reasonable range for a short period of time, within roughly two months ipify started experiencing issues once more. It’s growth had continued to rise at an astounding rate. Around this time I had set up Google Alerts for ipify, so that I’d know when people mentioned it. I started getting notifications that more and more people started using ipify in their personal and work projects. I then began receiving email from companies asking if they can embed it in their products (including a large smart TV provider, numerous media agencies, IoT vendors, etc.). Before I knew it, ipify was servicing around 15 billion requests/mo and I was now back to spending $50/mo on hosting costs.

I also started having issues with burst traffic – ipify would receive massive amounts of burst traffic for a short period of time that would die off quickly. I presumed this traffic was part of bootstrapping scripts, cron jobs, and other similar timed operations.

I later discovered through user notifications that antivirus vendors started blocking ipify because it started getting used in root kits, viruses, and other nasty pieces of software. Attackers would use ipify to get the victim’s public IP address before firing it off to a central location to be used for malicious purposes. I suppose this sort of usage was responsible for a large amount of that burst traffic.

While I’m not a fan of helping VX authors or spending money to assist them, I made the decision to keep ipify running neutrally to serve anyone who wants to use it. I’ve never been a fan of developer services that pick and choose what people use them for. I like to keep things simple.

Which left me to deal with the burst traffic problem. Dealing with burst traffic was tricky because I primarily had two choices:

  1. Run additional dynos at a cost to myself and be always prepared for burst traffic, or
  2. Use an auto-scaling service like AdeptScale to automatically create and destroy dynos based on traffic patterns on my behalf

I eventually decided to manually involve myself and go with option 1, simply because I didn’t want to spend any additional funds on another service (even though I’ve used AdeptScale’s service before, and it is fantastic). Around this time I was spending $150/mo and ipify was servicing ~25 billion requests/mo.

Which brings us to the recent past.

ipify Hits 30 Billion Requests per Month

Over the past few months, ipify has hit a new record and surpassed 30 billion requests/mo on several occasions. It was an exciting milestone to surpass and something that has been fun to watch.

Today, ipify routinely serves between 2k and 20k RPS (which is almost never consistent). Traffic is always variable and the usage is so routinely high that I’ve completely given up on trying to detect traffic patterns in any meaningful way. Average response time ranges between 1 and 20ms depending on traffic patterns.

Today the service runs for between $150/mo and $200/mo depending on burst traffic and other factors. If I factor that into my calculator (assuming a $200/mo spend), ipify is able to service each request for total cost of $0.000000007. That’s astoundingly low.

If you compare that to the expected cost of running the same service on a serverless or functions-as-a-service provider, ipify would cost thousands of dollars per month (as calculated here).

All in all, I’m extremely satisfied with ipify’s cost. It’s a frugal service that serves a simple purpose.

ipify's Future

Which leads me to the future. As ipify continues to grow, the service has gotten requests for a lot of different things: IPv6 support, better web design, other metadata about IP addresses, etc.

As I’m incredibly busy with other projects nowadays, I ended up passing ownership of ipify along to my good friends at https://www.whoisxmlapi.com.

Jonathan and team are good people working to build a portfolio of valuable and interesting developer API services.

While I still help out with things from time-to-time, Jonathan and team are currently implementing new features for ipify, and working hard to roll out some pretty cool changes that I’m excited about (including an improved UI and more data endpoint).

I look forward to seeing how ipify continues to grow over the coming years.

A version of this post was originally published on Randall Degges’ personal blog. And, while we might be biased, we'd be remiss if we didn’t recommend you check out Randall’s book, The Heroku Hacker’s Guide.

Originally published: January 23, 2018

Browse the archives for engineering or all blogs Subscribe to the RSS feed for engineering or all blogs.