Scalable Python Application Deployments on AWS OpsWorks
Here at Jazkarta we’ve been creating repeatable deployments for Plone and Django sites on Amazon’s AWS cloud platform for 6 years. Our deployment scripts were built on mr.awsome (a wrapper around boto, Python’s AWS library) and bash scripts. Our emphasis was on repeatable, not scalable (scaling had to be done manually), and our system was a bit cumbersome. There were aspects that were failure prone, and the initial deployment of any site included a fair amount of manual work. When we heard about AWS OpsWorks last summer we realized it would be a big improvement to our deployment process.
OpsWorks is a way to create repeatable, scalable, potentially multi-server deployments of applications on AWS. It is built on top of Chef, an open source configuration management tool. OpsWorks simplifies the many server configuration tasks necessary when deploying to the cloud; it allows you to create repeatable best practice deployments that you can modify over time without having to touch individual servers. It also lets you wire together multiple servers so that you can deploy a single app in a scalable way.
OpsWorks is a bit similar to PaaS offerings like Heroku but it is better suited for sites that need more scalability and customization than Heroku can provide. The cost of Heroku’s point and click simplicity is a lack of flexibility – OpsWorks lets you change things and add features that you need. And unlike the PaaS offerings, there is no charge for OpsWorks – you don’t pay for anything besides the AWS resources you consume.
Briefly, here’s how it works.
- First you create a stack for your application, which is a wrapper around everything. It may include custom chef recipes and configuration and it defines the AMI that will be used for all the instances. There are currently 2 choices, latest stable Ubuntu LTS or Amazon Linux. (We use Ubuntu exclusively.)
- Within the stack you define layers that represent the services or functionality your app requires. For example, you might define layers for an app server, front end, task queue, caching, etc. The layers define the resources they need – Elastic IPs, EBS volumes, RAID10 arrays, or whatever.
- Then you define the applications associated with the layers. This is your code, which can come from a github repo, an S3 bucket, or wherever.
- Then you define instances, which configure the EC2 instances themselves (size, availability zone, etc.), and assign the layers to the instances however you want. When you define an instance it is just a definition, it does not exist until it is started. Standard (24-hour) instances are started and stopped manually, time-based instances have defined start and stop days and times, and load-based instances have customizable CPU, load or memory thresholds which trigger instances to be started and stopped. When an instance starts, all the configuration for all the layers is run, and when it is stopped all the AWS resources associated with it are destroyed – aside from persistent EBS storage or Elastic IPs which are bound to the definition of the instance in OpsWorks instead of being bound to an actual instance.
For more details and a case study about switching from Heroku, see this excellent introduction to OpsWorks by the folks at Artsy.
What We Did
OpsWorks has native support for deploying Ruby on Rails, NodeJS, Java Tomcat, PHP, and static HTML websites, but no support for Python application servers (perhaps partly because there is no standard way to deploy Python apps). This was a situation we thought needed to be remedied. Furthermore, few if any PaaS providers are suitable for deploying the Plone CMS which many of our clients use. Because OpsWorks essentially allows you to build your own deployment platform using Chef recipes, it seemed like it might be a good fit.
Chef is a mature configuration management system in wide use, and there are many open source recipes available for deploying a variety of applications. None of those quite met our needs in terms of Python web application deployment, so we wrote two new cookbooks (a bundle Chef recipes and configuration). We tried to structure the recipes to mimic the built in OpsWorks application server layer cookbooks.
The repository is here: https://github.com/alecpm/opsworks-web-python. Each cookbook has its own documentation.
- Python Cookbook – provides recipes to create a Python environment in a virtualenv, to deploy a Django app, and to deploy a buildout
- Plone Cookbook – builds on the Python and buildout cookbooks to deploy scalable and highly available Plone sites
The Plone cookbook can handle complex Plone deployments. An example buildout is provided that supports the following layers:
- Clients and servers and their communication
- Load balancing
- Shared persistent storage for blobs
- Relstorage – either via Amazon RDS or a custom database server in its own OpsWorks layer (there is a built in layer for MySQL)
- Redis and Celery
- Auto-scaling the number of Zeo clients from the AWS instance size
The recipes handle automatically interconnecting these services whether they live on a single instance or on multiple instances across different Availability Zones. For more information, see the README in each cookbook.
We’ve used OpsWorks with our custom recipes on a few projects so far and are quite happy with the results. We have a wishlist of a few additional features that we’d like to add:
- Automated rolling deployments – a zero down time rolling deployment of new code that updates each Zeo client in sequence with a pause so the site doesn’t shut down.
- Native Solr support – use the OS packages to install Solr (instead of buildout and collective.recipe.solr) and allow custom configuration for use with alm.solrindex or collective.solr in the recipe.
- Integration of 3rd party (New Relic, Papertrail, …) and internal (munin, ganglia, …) monitoring services.
- Better documentation – we need feedback about what needs improvement.
If you’d like to contribute new features or fixes to the cookbooks feel free to fork and issue a pull request!