Anyone with the ability to merge PRs on GitHub can deploy the application. Whenever a PR is merged the code is deployed. When deploying complex code, be sure that other team members are around to help if something goes wrong.

Generally, it's a good idea to keep the SRE team in the loop on high risk deploys. However, deployments are our collective responsibility, so it's important to monitor your deploys. You can see deployment status on Travis-ci.com and in the #deployment-pipeline channel on Slack. Be prepared to rollback or push a fix for any deployment!

Forem relies on GitHub and Travis to deploy continuously to Heroku. If a Pull Request is merged it will automatically be deployed to production once the build steps complete successfully. The process currently takes about 20 minutes to complete and will need a few additional minutes before the change goes live.

The following stages can be explored in our .travis.yml and Procfile. Our Travis CI process consists of 2 stages.

1. Running our test suite in 3 parallel jobs.
2. Deploying the application.

Stage 1: Running Tests

In stage 1, we use KnapsackPro to divide our Rspec tests evenly between 3 different jobs (virtual machines). This ensures that each job takes relatively the same amount of time to run. After running our Rspec tests, we then run a series of other checks. These additional checks are split up between the different jobs. Here is a list of those additional checks that are run.

• Job 0 is where we run JavaScript tests, Preact tests, and coverage checks.
• Job 1 is where Travis builds Storybook to ensure its integrity, and where we check for any known vulnerabilities using bundle-audit.
• Job 2 is where Travis fires up a Rails console to ensure the application loads properly.

If all of the jobs pass then we move on to Stage 2 of the Travis CI process.

Stage 2: Deploying

If the build was kicked off from a pull request being created or updated this stage will do nothing. If the branch has been merged into master, then this stage will kick off a deploy. The deploy will run in its own job deploying our application to Heroku.

Prior to deploying the code, Heroku will run database migrations, Elasticsearch updates, and do some final checks (more information on that below) to make sure everything is working as expected. If these all succeed, then the deploy completes and our team is notified.

We use Heroku's Release Phase feature. Upon deploy, the app installs dependencies, bundles assets, and gets the app ready for launch. However, before it launches and releases the app Heroku runs a release script on a one-off dyno. If that release script/step succeeds the new app is released on all of the dynos. If that release script/step fails then the deploy is halted and we are notified.

The name of the script we use is release-tasks.sh and its in our root directory. During this release step we do a few checks.

1. We first check the DEPLOYSTATUS environment variable. In the event that we want to prevent deploys, for example after a rollback, we will set DEPLOYSTATUS to "blocked". This will cause the release script to exit with a code of 1 which will halt the deploy. This ensures that we don't accidentally push out code while we are waiting for a fix or running other tasks.
2. We run any outstanding migrations. This ensures that a migration finishes successfully before the code that uses it goes live.
3. We run any data update scripts that need to be run. A data update script is one that allows us to update data in the background separate from a migration. For example, if we add a new field to Elasticsearch and need to reindex all of our documents we would use a data update script.
4. We update Elasticsearch. Elasticsearch contains indexes which have mappings. Mappings are similar to database schema. The same way we run a migration to update our database we have to run a setup task to update any Elasticsearch mappings.
5. Following updating all of our datastores we use the Rails runner to output a simple string. Executing a Rails runner command ensures that we can boot up the entire app successfully before it is deployed. We deploy asynchronously, so the website is running the new code a few minutes after deploy. A new instance of Heroku Rails console will immediately run a new code.