Automatically creating a version of an application based on submitted pull request (PR) code and adding a link from the PR to the build can facilitate quicker turn around on feature iterations. This post describes one way to achieve this feature in your DevOps process by leveraging a new reverse-proxy server called Traefik, that can be configured to serve a Dockerized version of your application.
Traefik is a new powerful reverse proxy engine written in Golang. It can support all the traditional back-ends as well as newer ones and can do this dynamically. That is the major selling point for Traefik and it's a big one.
Traefik is currently about 85% as fast a Nginx (as reported on the Traefik site benchmark page). Considering how new this project is, and how much it can do, that performance is rather impressive.
Although Traefik is a new comer to the reverse proxy world, it has some relevant and exciting features that make it appropriate for our use case.
Now imagine if you were trying to set up a service to dynamically create a QA sites for your project:
Well, Traefik can be used to proxy to docker containers and can do this dynamically. This is huge because our use case will be to create QA sites for PRs using docker containers. Traefik can also be used on non container applications, including web or even flat files.
Conceptionally the requirements of this effort are easy to grok once we've settled on using Traefik with the docker back-end:
- First the web application we want to stage PRs for needs to be Dockerized (that's sort of a given really).
- Second, we need a way to interact dynamically with GitHub on events like PRs created and updated and merged, etc.. We can leverage the well known application Jenkins CI for this task since it's a stable project with plenty of GitHub web-hook plugins to choose from.
- Third we will need to do things like download a copy of the PR repo,
checkout PR branch or pull latest changes on branch, run
docker-compose up -dfrom within the PR repo source directory etc. We could of course do this inside of a Jenkins' job shell script, but that can get unwieldy to maintain, so instead I've encapsulated this concern into a gem created named Roper that I will use from withing the Jenkins job.
There are several options for setting up Traefik, including installing the binary. But, since we will be using Docker to deploy our apps anyway, it makes sense to also use Docker to deploy Traefik. A quick search on GitHub should lead you to a Dockerized setup for Traefik using the official Traefik docker image.
And that is precisely what I did. I found a project that sets up Traefik to
proxy a Dockerized back-end. I made some minor changes to the config file
Traefik.toml to fit my needs and spun it up:
git clone [email protected]:dkinzer/reverse-proxy.git
cd reverse-proxy.git
docker-compose up
And with that I have Traefik running as a service, and waiting for a docker-image to be spun up to serve on port 443.
Although Dockerizing your application is outside the scope of this post, there are some Traefik specific points that need to be kept in mind.
It's important to note that Traefik needs to belong to the same network as the back-end that it is serving. And if the back-end belongs to multiple networks then it needs to be configured so that it always communicates with Traefik using the network that they both share. This seems obvious in retrospect but I spent many hours trying to debug a situation where Traefik was randomly picking one network over another.
When using docker-compose to spin up your app, the way to configure what network Traefik listens to is via labels that can be passed configured.
In fact using these same labels you can configure whether or not Traefik will serve the Dockerized image at all (important when your app is composed of multiple services with one public facing service).
Below is an example docker-compose file that has been set up to work with Traefik:
services:
app:
build:
context: .
dockerfile: .docker/app/Dockerfile
container_name: "$ROPER_REPO_BRANCH"
ports:
- "3000"
environment:
SOLR_URL: "http://solr:8983/solr/blacklight"
LC_ALL: "C.UTF-8"
labels:
- "traefik.enable=true"
- "traefik.frontend.rule=Host:${ROPER_REPO_BRANCH}.${DOMAIN}"
- "traefik.backend=microbot"
- "traefik.docker.network=reverseproxy_default"
networks:
- "reverseproxy_default"
depends_on:
- solr
entrypoint:
- ruby
- bin/app-start.rbNote that in the section above we can set up even the front-end rules for what the service Host will be. We'll talk more about the environment variables being set in that line in the section below on the Roper gem.
You can learn more about what labels are available for configuring a Docker back-end at: https://docs.traefik.io/configuration/backends/docker/
Since the goal is to get something to happen when a GitHub PR is generated, then we can use a GitHub web-hook to post a Jenkins Job: https://developer.github.com/webhooks/. The Jenkins Job in turn could trigger the setting up of containerized application.
For our particular case we decided to use two Jenkins Jobs. One job is to handle the web-hook and the other one is a parameterized job to actually trigger the build. The reason for this two step process is that with the parameterized build we can run rebuilds in the cases where a build goes wrong (or just because), and that is a nice feature to have.
For the sake of brevity we'll paraphrase each of the jobs below with links to the full job definitions:
This job leverages the Jenkins Generic Trigger plugin. It uses JSONPaths to parse the GitHub web-hook payload:
| key | JSONPath |
|---|---|
| repo | $.pull_request.head.repo.full_name |
| branch | $.pull_request.head.ref |
| sha | $.pull_request.head.sha |
| action | $.action |
We then define the following filter in order to only trigger on specific matches.
| filter | match |
|---|---|
$X_GitHub_Event $repo $action |
pull_request tulibraries\/tul_cob (open|close|synchronize) |
Thus we limit the build triggers to pull requests coming from PRs generated from a specific organization and repository.
For the build step we use the parameters we collected to trigger a parameterized job.
The receiving job is a parameterized job that simply calls out to the actual building process.
args="--repo=$repo --branch=$branch"
if [ "$action" == "closed" ]; then
command=release
else
command=lasso
args=";$args --domain=devkinzer.net --sha=$sha --status_url=${BUILD_URL}console";
fi
ssh devkinzer -C roper $command $argsSo as you can see we are calling out to an executable named roper and
specifying arguments for a build.
Once we have created a Jenkins Job that we can trigger via a web-hook post then its time to configure the GitHub web-hook. There are instructions on GitHub of how to go about this so I wont go into any details here. The important bit for us is to before sure that:
- The web-hook is set to have content type application/json
- And, it should trigger at least for PRs and push events.
Of course instead of calling out to a build executable we could have written a build script directly into a Jenkins job. Such a script would take care of all the steps for properly launching the app such as:
- Posting to the GitHub PR a status of change
- Cloning the PR and checking out the correct reference
- Running
docker-compose upwithin the repository's working directory. - Posting a link to the built image that gets served by Traefik
- Handling updates to the PR properly
- Handling PR closing such as removing the images and files used in the build.
Although this list is small an relatively straight forward. Such a script could easily become unwieldy to maintain as a Jenkins Job shell step. Instead I chose to wrap this functionality into a gem which I named Roper. As a gem I can apply tests to all of the features and separate out concerns to make this easier to understand user and maintain in the long run.
In addition to the steps above Roper adds some environment variables that are available to docker-compose. Thus in the example above:
- "traefik.frontend.rule=Host:${ROPER_REPO_BRANCH}.${DOMAIN}"
ROPER_REPO_BRANCH is made available by Roper. For more details on Roper see
the Roper documentation
I want to end this post by touching a little on another incentive for this project and also touch on some of the drawbacks of this particular solution.
So I've shown one way that can be used to generate sites based on PRs. I mentioned in the abstract that one incentive for implementing such a feature is to facilitate quicker turn around on feature iterations. In a nutshell, the idea here is that we lower the bar for being able to review a PR by automatically building a version of the app based on the PR. Thus, for example, a developer would not have to take time to download and build the PR locally to review some aspects of the PR. But this same feature can be used by a non developer. In other words, we can make it easier for non developers to be involved earlier in the feature iteration process. This is of course a very important incentive.
Another incentive for this project, as I see it, is that it's a cost effective alternative to similar commercial solutions. (Well, at least one other solution I know of which cost a previous employer of mine $300 per month).
That said, what is the cost of this solution? Well that will depend on the need and how much resources your app requires. By need I mean the calculus of the max number of PRs you expect to be opened multiplied by the requirements of your apps. So, for example I have an app that requires about 800MB to run. And on a small Linode VM, I handle about 4 PRs before the little machine starts running out of memory and breaking.
And that also gets at one of the drawbacks of the process as I describe it here. Since (depending on how much space and memory your app requires) the resources of the machine that is providing the service can be quickly exhausted under certain stresses.
But considering that you may pay up to $300/month for a similar feature using a vendor solution, it's likely you can provision a VM node with plenty of resources for less than that.
That said, it's important to note that Traefik can also be set on other back-ends that can potentially scale for just about any use case.