Cloudinary Blog

A story about production systems, Rails, monitoring and off-hour notifications

Cloudinary's image management service is used by thousands of world-wide websites and mobile apps. For many of our clients, Cloudinary has become a central, mission-critical component used for managing image uploads, transformations and delivery.
 
This is why we've built Cloudinary from the ground up to be a very robust service. We put a lot of emphasis on availability, scalability and support and we take our users' confidence in us extremely seriously.
 
So far, we've been quite satisfied with our ability to keep Cloudinary at an average of > 99.99% uptime.
 
However, on April 4th, the Cloudinary service experienced outages for a few hours. We wanted to explain what happened, our conclusions and the steps we've taken to make sure this won’t happen again.
 

The upgrade

Cloudinary's core service is built with Ruby on Rails. The service is tested thoroughly and upgrades are handled with uttermost care. This is why we've preferred to stay with Rails v3.0 for a long time rather than rock the boat with an upgrade to the latest Rails 3.2.
 
A few weeks ago a security vulnerability was discovered in Rails. As always, we wanted to apply the security fix as soon as possible. However, the Rails team stopped releasing fixes for Rails 3.0. We had to upgrade to v3.2.
 
We've upgraded to Rails 3.2 in our lab and modified our code to support it (Rails upgrades tend to be non backward compatible and break code built with previous versions). We've tested our code extensively and verified that our thousands of unit tests passed correctly. We've successfully finished a thorough manual QA of the system in our staging environment. It all went quite smoothly.
 
We scheduled the upgrade for April 4th. As usual, we deployed the system gradually to all of our production servers. Deployment went smoothly as well. We performed additional sanity testing after the system was deployed and closely monitored the system during the working day.
 
We went to sleep happy and relaxed.
 

The issues

At about 1am at night things started to shake.
 
Apparently, Rails 3.2 changed the defaults of one simple configuration parameter - response caching was turned on by default when certain cache headers are returned.
 
As a result, after long hours of service requests, the local application disk for some of our servers became full due to the cached responses. This caused certain requests that required disk space to fail, depending on the exact request and the size of the response.
 
Annoyingly enough, the automatic monitoring service that regularly verifies our APIs, was performing a request that required very little disk space and continued to operate regularly. This service is configured to send notifications to our engineering team's mobile phones during the night. But since no errors were detected, no notification was sent.
 
Fortunately for us, our co-founder's toddler woke him up early in the morning. He naturally (?) checked his inbox, understood that something was very wrong. He quickly cleared the disk space and modified Rails 3.2 cache settings. The system was fully working again.
 
It's important to note that during these ~5 hours, all existing images and transformed images were delivered successfully to users through our delivery service and tens of thousands of worldwide CDN edges (Akamai + CloudFront). Still, part of the upload API calls did fail during this time and we are very sorry for this.
 

Going forward

Naturally, we've immediately started to improve our outage prevention mechanisms.
 
We've added additional disk space tests to our QA list and added abnormal disk usage monitoring to our urgent notification service. We're also adding a wider set of API requests to our automatic service monitoring.
 
We've integrated with Twilio to enhance our off-hour notifications. Specifically, our engineering team will now receive automatic voice calls to their mobile phones in addition to our previous notification methods.
 
To make sure we keep you in the know during outages, we've pushed up the priority of a public status page. This page will include automatic monitoring details as well as human written notes.
 

Summary & conclusions 

We are happy that Cloudinary had nearly zero availability issues in almost 2 years of operations. On the other hand, no online service is perfect and every service experienced or will experience outages. 
 
We will continue to enhance our service with additional image-related features. On the same time, we'll continue to work hard on having Cloudinary's uptime as close to 100% as possible.
 
Thank you for trusting us with your images!

Recent Blog Posts

React.js Tutorial: How to develop a React library

Developing a library requires a different approach from developing an application. You must consider the use of the library in someone else’s application and design for it. React is well suited for this purpose. And if the library you are creating is an adapter to another library, you can dynamically generate the component's properties definition to ensure they are forward compatible. There is however more than one way to achieve the same goal, with some conventions to follow and others to cautiously not follow. In particular, I chose to use the context function even though it is an experimental feature because it is useful when you don’t know, or can’t dictate, the way your library's components will be utilized.

Read more
Learn about AdonisJs and how to set up a CMS with it

Even though Node is fun, easy and cheap to work with, we spend a lot of time writing boilerplate codes because structure and organization is missing.

What happened to Convention Over Configuration?

While Node is simple, it requires you to make a lot of decisions, which ultimately causes confusion because it leaves you with several options. Languages like PHP, Ruby, C# and Python have one or more Molde-View-Controller (MVC) frameworks, such as Laravel, Rails, ASP.Net and Django. These help developers to achieve structure and write maintainable code with these languages. That was not the case for Node until AdonisJs was introduced.

Read more
Cloud-based image filter transformations for developers

Every picture has a story to tell. But the story it tells can change when you change the color tone, saturation, contrast, or other elements of a photo.

A few years ago, post-processing a digital image generally required a high level of skill and expensive software such as PhotoShop. But in recent years, popular photo sharing apps such as Instagram, Flickr, and Snapchat started offering built-in filters. Professionals take advantage of filters to make subtle corrections or adjustments. Casual users often apply more prominent filters that add their own unique touch or just make their images more fun.

Read more
Optimized media delivery with multi-CDN solutions
This article originally appeared on Venture Beat Magazine and is reprinted with permission.
 
As we all know, content that takes forever to load will increase your site’s bounce rates, decrease your conversions, and undeniably scar your bottom line. And by “forever” I mean more than a few seconds. A recent study by KISS Metrics found that page abandonment rates increase drastically after a load time of just four seconds or more.
Read more