OEP-5: Pre-built Developer Environments#
Pre-built Developer Environments
Calen Pennington <firstname.lastname@example.org>
Jesse Zoldak <email@example.com>
This OEP has been replaced by Tutor as a replacement for edx/configuration a decision made as a part of OEP-45: Configuring and Operating Open edX
Open edX is a large ecosystem of smaller services. In order to facilitate newcomers working in that system, each service will provide a pre-built environment that is packaged, published, and ready to be used for development.
Open edX and the edx-platform repository have gone through several iterations of developer setup instructions.
The first set of instructions detailed how to install edx-platform and all of its dependencies on your local workstation. As the number of developers and developer environments increased, so did the workload of maintaining instructions to work around all of the idiosyncrasies in those environments. These instructions weren’t useful for production installations, which were standardized on a particular release of the Ubuntu operating system.
The second, current, method of development is referred to as
is a Vagrant image that comes prepackaged with installations of
edx-platform, forums, and an increasingly large number of other services, as
well as single shared instances of the databases that they all need to run.
This marked a significant improvement over the ad-hoc installation methods,
because it was able to use the infrastructure that was built out for installing
Open edX in production. However, there are several problems with the current
monolithic Vagrant image which we would like to address:
As new services are added to the Open edX ecosystem, the Vagrant image grows in size, requiring more system memory from the host machine.
Vagrant imposes noticeable overhead during file system access, which especially hurts any development that is doing file-heavy operations (such as manipulating static assets).
The monolithic Vagrant image isn’t particularly representative of the production service architecture, because all of the services are installed on the same machine, which makes it harder for developers to be confident that their changes will work when deployed.
The shared services dependencies on things like MySQL and Elasticsearch mean that all of the services in Devstack have to be upgraded at the same time, or have to maintain compatibility with multiple database versions until all services are ready to update. This hinders independent development on the services.
Because all of the services are installed on the same image, the image either has to be updated by every user after downloading (to update to the latest code versions), or the image has to be rebuilt every time any of the installed services change (which results in lots of churn).
Because the Vagrant image contains many separate services, the definition of that image has to be separate from any of the service repositories, which makes it hard to update (and to push updates out to users).
The primary developer interaction with the local development environment will
make shortcuts in the repository containing the code for the
service being developed. For example, in order to work on an
local development environment, a developer could run
make dev.watch from a
checked-out edx-platform repository.
The following targets will be standard:
make dev.watch # Start the local development environment services (if # they aren't already) and display their logs on stdout. make dev.shell # Start the local development environment services (if they # aren't already) and access a shell on the primary # container (as the user that owns the service, in the # directory that has the service source code). make dev.clean # Stop the local development environment services and # remove any intermediate data that they stored in the # .dev/ directory. make dev.* # Start the local development environment services (if they # aren't already) and run `make *` from the code directory # of the primary service (as the service owner).
Third-party services like MySQL and MongoDB should be run using vendor provided official containers. If no reasonable alternative exists to customizing a provided container, the minimal changes should be applied as a layer on the provided container.
Rationale: We should be aiming to minimize the amount of work needed to support our development environments. Minimizing changes to third-party containers means we have less code/configuration that we need to update as third-party containers and applications change. Minimizing customization should also push us in the direction of making it easy to deploy our stack, because there will be fewer customizations needed by operators of Open edX in production.
Open edX containers must run all of their processes via supervisor to provide a single entry point even when the container has multiple processes (such as nginx and a Django service running under gunicorn).
Rationale: Docker imposes a single-entry-point constraint. Standardizing on supervisor means that we need to make fewer choices during development, and that we can combine the learnings and configurations across various Open edX projects.
Open edX services running in containers must log to
Rationale: Logging to
stderrmakes it significantly easier to develop the app in different environments, since those file handles are available nearly everywhere. Supervisor captures both file handles and logs their contents to disk, which means that we can display the logs to the screen, and can search through old development logs.
Dependent services (such as databases) should be configured by default to mount their storage directories inside
.dev/inside the source code directory (this should be added to .gitignore and .dockerignore). This location should be configurable via an environment variable.
Rationale: This will allow data from local development environment to persist across container executions. Locating the storage in the source code repository isolates the storage of third-party containers that are used in common between local development environments so that data doesn’t leak between them. Making the storage location configurable means that local development environments could share storage if desired.
The Open edX service must mount the source code directory from the host machine into the development container so that the local development environment is always running the version of code being edited.
Rationale: The development container should always run the latest source code that the developer is editing. Having it mount the local source code means that the developer can edit on their local machine and have their changes reflected immediately in the container.
The local development environment should be configured to auto-reload when source code is modified.
Rationale: Fast feedback loops are key to developer productivity. Having the application reflect changes made by the developer as soon as possible allows the developer to test and validate their changes quickly.
Open edX Service containers should be built during CI, and published so that any developer can download the latest container. They should also be tagged for each versioned release.
Rationale: Having a container that is up-to-date with both code and configuration makes it easier for developers to make their changes with confidence that those same changes will work once they are merged. Having containers for previous releases allows easier development for hot fixes.
The local development environment should be installed with all pre-requisites to allow tests to be run on them without additional developer input.
Rationale: Tests are an important part of the fast feedback loop for developers. Being able to run those tests immediately when the local development environment starts means that developers can have confidence that the environment is working correctly for development.
Development containers should be designed for development over production.
Rationale: The local development environment is intended to make it easy for new and existing developers to make changes to Open edX services. As such, they should be focussed first on development, and only secondarily on production-readiness or production-fidelity.
The prototype implementation was done using Docker for the containers, and Docker Compose to manage the networking between services.
Docker has a low per-service overhead, because it doesn’t allocate resources to particular containers.
Docker uses a layered file system, which, if used properly, should let us limit the amount of data that a developer needs to download to update from one version of the base image to the next.
This implementation supersedes and replaces use of Vagrant for the local development environment. However, nothing in the implementation should break the Vagrant-based local development environment during the development phase, and developers currently using Vagrant will be able to continue to use it. Long-term, the goal will be to deprecate and no longer support the Vagrant-based local development environment.
The course-discovery service uses a Docker-based local development environment, and was the prototype implementation of this OEP. However, it will need modification to fully meet the specifications of this OEP.