runboat ☸️

A simple Odoo runbot lookalike on kubernetes. Main goal is replacing the OCA runbot.

Principle of operation

This program is a Kubernetes operator that manages Odoo instances with pre-installed addons. The addons come from commits on branches and pull requests in GitHub repositories. A deployment of a given commit of a given branch or pull request of a given repository is known as a build.

Runboat has the following main components:

• An in-memory database of deployed builds, with their current status.

• A REST API to list builds and trigger new deployments as well as start, stop, redeploy or undeploy builds.

• A GitHub webhook to automatically trigger new builds on pushes to branches and pull requests of supported repositories.

• A controller that performs the following tasks:

  • monitor deployments in a kubernetes namespaces to maintain the in-memory database;
  • on new deployments, trigger an initialization job to create the corresponding postgres database and install the addons in it;
  • initialization jobs are started concurrently up to a configured limit;
  • when the initialization job succeeds, scale up the deployment, so it becomes accessible;
  • when the initializaiton job fails, flag the deployment as failed;
  • when there are too many deployments started, stop the oldest started;
  • when there are too many deployments, delete the oldest created;
  • when a deployment is deleted, run a cleanp job to drop the database and delete all kubernetes resources associated with the deployment.

When a deployment is stopped, the corresponding postgres database remains present, so deployments can restart almost instantly.

This approach allows the deployment of a very large number of builds which consume no memory nor CPU until they are started. The number of started deployment can also be high, by reserving limited CPU and memory resources for each, taking advantage of the fact that they are typically used infrequently. The number of concurrent initialization jobs is limited strongly, and they are queued, as these are typically the more resource-intensive part of the lifecycle of builds.

All state is stored in kubernetes resources (labels and annotations on deployments). The controller can be stopped and restarted without losing state.

Requirements

For running the builds:

• A namespace in a kubernetes cluster.
• A wildcard DNS domain that points to the kubernetes ingress.
• A postgres database, accessible from within the cluster namespace with a user with permissions to create database.

For running the controller (runboat itself):

• Python 3.10
• kubectl
• A KUBECONFIG or an in-cluster service account that provides access to the namespace where the builds are deployed, with permissions to create and delete Service, Job, Deployment, Ingress, Secret and ConfigMap resources as well as read and watch Deployments and Jobs.
• Some sort of reverse proxy to expose the REST API.

The controller can be run outside the kubernetes cluster or deployed inside it, or even in a different cluster.

Developing

• setup environment variables (start from .env.sample, the meaning of the environment variables is documented in settings.py)
• create a virtualenv, make sure to have pip>=21.3.1 and pip install -c requirements.txt -e .[test]
• run with uvicorn runboat.app:app --log-config=log-config.yaml
• api documentation is at http://localhost:8000/docs
• run tests with pytest (environment variables used in tests are declared in .env.test)

Running in production

gunicorn -w 1 -k runboat.uvicorn.RunboatUvicornWorker runboat.app:app.

One and only one worker process !

Gunicorn also necessary so SIGINT/SIGTERM shutdowns after a few seconds. Since we use run_in_executor, SIGINT/SIGTERM handling does not work very well, and gunicorn makes it more robust. https://bugs.python.org/issue29309

Kubernetes resources

All resources to be deployed in kubernetes for a build are in src/runboat/kubefiles. They are gathered together from a kustomization.yaml jinja template that leads to three possible resource groups depending on a mode variable in the jinja rendering context:

• deployment creates a kubernetes deployment with its associated service and ingress;
• initialization creates a job that creates the database;
• cleanup creates a job that drops the database;

Besides the three modes, the controller has limited knowledge of what the kubefiles actually deploy. It expects the following to hold true:

• the runboat/build label is set on all resources, with the unique build name as value;

• a deployment starts with 0 replicas and is created with a runboat/init-status=todo label, as well as a runboat/cleanup finalizer;

• the intialization job and pods have a runboat/job-kind=initialize label;

• the cleanup job and pods have a runboat/job-kind=cleanup label.

• the following annotations are set on deployments:

  • runboat/repo: the repository in owner/repo format;
  • runboat/target-branch: the branch or pull request target branch;
  • runboat/pr: the pull request number if this build is for a pull request;
  • runboat/git-commit: the commit sha.

• the home page of a running build is exposed at http://{build_slug}.{build_domain}.

During the lifecycle of a build, the controller does the following on the deployed resources:

• it sets the runboat/init-status annotation (todo, started, succeeded, failed) on deployments to track the outcome of the initialization jobs;
• it sets the deployment's specs.replica to 1 or 0 to start or stop it;
• it deletes the deployment when an undeploy is requested (the actual delete occurs later due to the finalizer);
• it removes the deployment finalizers and deletes resources matching the runboat/build label after the cleanup job succeeded.

TODO

Advanced prototype (min required to open the project):

• plug it on a bunch of OCA and shopinvader repos to test load

MVP:

• better error handling in API (return 400 on user errors)
• more tests
• look at other TODO in code to see if anything important remains
• basic UI (single page with a combo box to select repo and show builds by branch/pr, with start/stop buttons)
• secure github webhooks
• deployment and more load testing

More:

• streaming build/log and build/init-log api endpoints
• shiny UI
• websocket stream of build changes, for a dynamic UI
• handle PR close (delete all builds for PR)
• handle branch delete (delete all builds for branch)
• create builds for all supported repos on startup (goes with sticky branches)
• never undeploy last build of sticky branches
• configurable kubefiles directory
• even more tests

Author and contributors

Authored by Stéphane Bidoul (@sbidoul).

Contributions welcome.