Bike Spike: Scaling Cascade Bicycle Club’s Web Application

As the largest bicycling club in the country with more than 16,000 active members and a substantially larger community across the Puget Sound, Cascade Bicycle Club requires serious performance from its website. For most of the year, Cascade.org serves a modest number of web users as it furthers the organization’s mission of “improving lives through bicycling.”

But a few days each year, Cascade opens registration for its major sponsored rides, which results in a series of massive spikes in traffic. Cascade.org has in the past struggled to keep up with demand during these spikes. During the 2014 registration period for example, site traffic peaked at 1,022 concurrent users and >1,000 transactions processed within an hour. The site stayed up, but the single web server seriously struggled to stay on its feet.

In preparation for this year’s event registrations, we implemented horizontal scaling at the web server level as the next logical step forward in keeping pace with Cascade’s members. What is horizontal scaling, you might ask? Let me explain.

[Ed Note: This post gets very technical, very quickly.]

Overview

We had already set up hosting for the site in the Amazon cloud, so our job was to build out the new architecture there, including new Amazon Machine Images (AMIs) along with an Autoscale Group and Scaling Policies.

Here is a diagram of the architecture we ended up with. I’ll touch on most of these pieces below.

Web Servers as Cattle, Not Pets

I’m not the biggest fan of this metaphor, but it’s catchy: The fundamental mental shift when moving to automatic scaling is to stop thinking of the servers as named and coddled pets, but rather as identical and ephemeral cogs–a herd of cattle, if you will.

In our case, multiple web server instances are running at a given time, and more may be added or taken away automatically at any given time. We don’t know their IP addresses or hostnames without looking them up (which we can do either via the AWS console, or via AWS CLI — a very handy tool for managing AWS services from the command line).

The load balancer is configured to enable connection draining. When the autoscaling group triggers an instance removal, the load balancer will stop sending new traffic, but will finish serving any requests in progress before the instance is destroyed. This, coupled with sticky sessions, helps alleviate concerns about disrupting transactions in progress.

The AMI for the “cattle” web servers (3) is similar to our old single-server configuration, running Nginx and PHP tuned for Drupal. It’s actually a bit smaller of an instance size than the old server, though — since additional servers are automatically thrown into the application as needed based on load on the existing servers — and has some additional configuration that I’ll discuss below.

As you can see in the diagram, we still have many “pets” too. In addition to the surrounding infrastructure like our code repository (8) and continuous integration (7) servers, at AWS we have a “utility” server (9) used for hosting our development environment and some of our supporting scripts, as well as a single RDS instance (4) and a single EC2 instance used as a Memcache and Solr server (6). We also have an S3 instance for managing our static files (5) — more on that later.

Handling Mail

One potential whammy we caught late in the process was handling mail sent from the application. Since the IP of the given web server instance from which mail is sent will not match the SPF record for the domain (IP addresses authorized to send mail), the mail could be flagged as spam or mail from the domain could be blacklisted.

We were already running Mandrill for Drupal’s transactional mail, so to avoid this problem, we configured our web server AMI to have Postfix route all mail through the Mandrill service. Amazon Simple Email Service could also have been used for this purpose.

Static File Management

With our infrastructure in place, the main change at the application level is the way Drupal interacts with the file system. With multiple web servers, we can no longer read and write from the local file system for managing static files like images and other assets uploaded by site editors. A content delivery network or networked file system share lets us offload static files from the local file system to a centralized resource.

In our case, we used Drupal’s S3 File System module to manage our static files in an Amazon S3 bucket. S3FS adds a new “Amazon Simple Storage Service” file system option and stream wrapper. Core and contributed modules, as well as file fields, are configured to use this file system. The AWS CLI provided an easy way to initially transfer static files to the S3 bucket, and iteratively synch new files to the bucket as we tested and proceeded towards launch of the new system.

In addition to static files, special care has to be taken with aggregated CSS and Javascript files. Drupal’s core aggregation can’t be used, as it will write the aggregated files to the local file system. Options (which we’re still investigating) include a combination of contributed modules (Advanced CSS/JS Aggregation + CDN seems like it might do the trick), or Grunt tasks to do the aggregation outside of Drupal during application build (as described in Justin Slattery’s excellent write-up).

In the case of Cascade, we also had to deal with complications from CiviCRM, which stubbornly wants to write to the local file system. Thankfully, these are primarily cache files that Civi doesn’t mind duplicating across webservers.

Drush & Cron

We want a stable, centralized host from which to run cron jobs (which we obviously don’t want to execute on each server) and Drush commands, so one of our “pets” is a small EC2 instance that we maintain for this purpose, along with a few other administrative tasks.

Drush commands can be run against the application from anywhere via Drush aliases, which requires knowing the hostname of one of the running server instances. This can be achieved most easily by using AWS CLI. Something like the bash command below will return the running instances (where ‘webpool’ is an arbitrary tag assigned to our autoscaling group):

[cascade@dev ~]$aws ec2 describe-instances --filters "Name=tag-key, Values=webpool" |grep ^INSTANCE |awk '{print $14}'|grep 'compute.amazonaws.com'

We wrote a simple bash script, update-alias.sh, to update the ‘remote-host’ value in our Drush alias file with the hostname of the last running server instance.

Our cron jobs execute update-alias.sh, and then the application (both Drupal and CiviCRM) cron jobs.

Deployment and Scaling Workflows

Our webserver AMI includes a script, bootstraph.sh, that either builds the application from scratch — cloning the code repository, creating placeholder directories, symlinking to environment-specific settings files — or updates the application if it already exists — updating the code repository and doing some cleanup.

A separate script, deploy-to-autoscale.sh, collects all of the running instances similar to update-alias.sh as described above, and executes bootstrap.sh on each instance.

With those two utilities, our continuous integration/deployment process is straightforward. When code changes are pushed to our Git repository, we trigger a job on our Jenkins server that essentially just executes deploy-to-autoscale.sh. We run update-alias.sh to update our Drush alias, clear the application cache via Drush, tag our repository with the Jenkins build ID, and we’re done.

For the autoscaling itself, our current policy is to spin up two new server instances when CPU utilization across the pool of instances reaches 75% for 90 seconds or more. New server instances simply run bootstrap.sh to provision the application before they’re added to the webserver pool.

There’s a 300-second grace time between additional autoscale operations to prevent a stampede of new cattle. Machines are destroyed when CPU usage falls beneath 20% across the pool. They’re removed one at a time for a more gradual decrease in capacity than the swift ramp-up that fits the profile of traffic.

More Butts on Bikes

With this new architecture, we’ve taken a huge step toward one of Cascade’s overarching goals: getting “more butts on bikes”! We’re still tuning and tweaking a bit, but the application has handled this year’s registration period flawlessly so far, and Cascade is confident in its ability to handle the expected — and unexpected — traffic spikes in the future.

Our performant web application for Cascade Bicycle Club means an easier registration process, leaving them to focus on what really matters: improving lives through bicycling.