Previously, configuration management was very confusing because we kept mixing
"base" and "defaults" configuration:
- It was difficult to make the difference between core settings that were
necessary (e.g: passwords) as opposed to others that could simply be
defaulted to.
- The order of settings in config.yml mattered: config entries that depended on
other needed to be defined later. As a consequence, Tutor was not compatible
with Python 3.5, where dict entries are not sorted.
When disable a plugin that set config entried, such as the minio plugin, tutor was logging the following:
Disabling plugin minio...
Removed config entry OPENEDX_AWS_ACCESS_KEY=openedx
Removed config entry OPENEDX_AWS_SECRET_ACCESS_KEY={{ MINIO_AWS_SECRET_ACCESS_KEY }}
Plugin disabled
The config values were not rendered during printing, which is a shame, because
the whole point of this log line is to warn users of passwords/secrets that are
being removed. Here, we make sure that the config values are properly rendered.
The new logs are now:
Disabling plugin minio...
Removing config entry OPENEDX_AWS_ACCESS_KEY=openedx
Removing config entry OPENEDX_AWS_SECRET_ACCESS_KEY=64vpCVLxhDxBuNjakSrX4CQg
Plugin disabled
We were forced to pin click to < v8 because of missing dependencies. In
particular, click_repl was broken. This is no longer the case, as click_repl
0.20 was published. Also, Jinja2 now includes type annotations, which allows us
to get rid of a few "# type: ignore" statements.
We take the opportunity to upgrade all requirements, which allows us resolve a
security issue on urllib3<1.26.0.
When running `tutor local quickstart -p` we were getting the following error:
Usage: custom [OPTIONS] ARGS...
Try 'custom --help' for help.
Error: Missing argument 'ARGS...'.
The docker-compose command sometimes accept a single command ("pull") with zero
argument.
See: https://discuss.overhang.io/t/local-quickstart-not-working-when-pullimages-enabled/1526
I stumbled upon a bug that should have been detected by the type
checking. Turns out, considering that config is of type Dict[str, Any]
means that we can use just any method on all config values -- which is
terrible. I discovered this after I set `config["PLUGINS"] = None`:
this triggered a crash when I enabled a plugin.
We resolve this by making the Config type more explicit. We also take
the opportunity to remove a few cast statements.
When the PLUGINS config entry is None (`PLUGINS:`), the following error
was being triggered:
File "/.../tutor/tutor/plugins.py",
line 304, in is_enabled
return name in config.get(CONFIG_KEY, [])
TypeError: argument of type 'NoneType' is not iterable
Annotations were generated with pyannotate:
https://github.com/dropbox/pyannotate
We are running in strict mode, which is awesome!
This affects a large part of the code base, which might be an issue for
people running a fork of Tutor. Nonetheless, the behavior should not be
affected. If anything, this process has helped find and resolve a few
type-related bugs. Thus, this is not considered as a breaking change.
This introduces a new dev/local command:
tutor dev bindmount CONTAINER PATH
And a new volume syntax:
tutor dev run --volume=PATH CONTAINER
This syntax automatically bind-mounts folders from the tutorroot/volumes
directory, which is pretty nifty.
- 💥[Improvement] Upgrade Open edX to Koa
- 💥 Setting changes:
- The ``ACTIVATE_HTTPS`` setting was renamed to ``ENABLE_HTTPS``.
- Other ``ACTIVATE_*`` variables were all renamed to ``RUN_*``.
- The ``WEB_PROXY`` setting was removed and ``RUN_CADDY`` was added.
- The ``NGINX_HTTPS_PORT`` setting is deprecated.
- Architectural changes:
- Use Caddy as a web proxy for automated SSL/TLS certificate generation:
- Nginx no longer listens to port 443 for https traffic
- The Caddy configuration file comes with a new ``caddyfile`` patch for much simpler SSL/TLS management.
- Configuration files for web proxies are no longer provided.
- Kubernetes deployment no longer requires setting up a custom Ingress resource or custom manager.
- Gunicorn and Whitenoise are replaced by uwsgi: this increases boostrap performance and makes it no longer necessary to mount media folders in the Nginx container.
- Replace memcached and rabbitmq by redis.
- Additional features:
- Make it possible to disable all plugins at once with ``plugins disable all``.
- Add ``tutor k8s wait`` command to wait for a pod to become ready
- Faster, more reliable static assets with local memory caching
- Deprecation: proxy files for Apache and Nginx are no longer provided out of the box.
- Removed plugin `{{ patch (...) }}` statements:
- "https-create", "k8s-ingress-rules", "k8s-ingress-tls-hosts": these are no longer necessary. Instead, declare your app in the "caddyfile" patch.
- "local-docker-compose-nginx-volumes": this patch was primarily used to serve media assets. The recommended is now to serve assets with uwsgi.
Tutor was making many calls to iter_installed (~100 on my machine with a
dozen installed plugins). Turns out it's useless to cache Plugin and
Renderer instances, as the config keeps changing all the time. Instead,
we cache the list of installed plugins, which does not change in the
course of a single run.
On my machine this speeds up `tutor config save` by 5x, going from 7.5s
to 1.3s.
Here, we upgrade the Open edX platform from Ironwood to Juniper. This
upgrade does not come with many feature changes, but there are many
technical improvements under the hood:
- Upgrade from Python 2.7 to 3.5
- Upgrade from Mongodb v3.2 to v3.6
- Upgrade Ruby to 2.5.7
We took the opportunity to completely rething the way locally running
platforms should be accessed for testing purposes. It is no longer
possible to access a running platform from http://localhost and
http://studio.localhost. Instead, users should access
http://local.overhang.io and https://studio.local.overhang.io. This
drastically simplifies internal communication between Docker containers.
To upgrade, users should simply run:
tutor local quickstart
For Kubernetes platform, the upgrade process is outlined when running:
tutor k8s upgrade --from=ironwood
Running jobs was previously done with "exec". This was because it
allowed us to avoid copying too much container specification information
from the docker-compose/deployments files to the jobs files. However,
this was limiting:
- In order to run a job, the corresponding container had to be running.
This was particularly painful in Kubernetes, where containers are
crashing as long as migrations are not correctly run.
- Containers in which we need to run jobs needed to be present in the
docker-compose/deployments files. This is unnecessary, for example when
mysql is disabled, or in the case of the certbot container.
Now, we create dedicated jobs files, both for local and k8s deployment.
This introduces a little redundancy, but not too much. Note that
dependent containers are not listed in the docker-compose.jobs.yml file,
so an actual platform is still supposed to be running when we launch the
jobs.
This also introduces a subtle change: now, jobs go through the container
entrypoint prior to running. This is probably a good thing, as it will
avoid forgetting about incorrect environment variables.
In k8s, we find ourselves interacting way too much with the kubectl
utility. Parsing output from the CLI is a pain. So we need to switch to
the native kubernetes client library.
Clarify a few variable names, make code more modular. Also, the Renderer
class now makes more sense as a singleton. We took the opportunity to
delete quite a lot of code.
This is for supporting json-based plugins. The great thing about this
change is that it allows us to easily print plugin version numbers in
`plugins list`.
The `dev` commands now rely on a different openedx-dev docker image.
This gives us multiple improvements:
- no more chown in base image
- faster chown in development
- mounted requirements volume in development
- fix static assets issues
- bundled ipdb/vim/... packages, which are convenient for development
Close #235
This has an impact on plugin hooks. Plugin hooks that needed to run
inside mysql-client now need to run inside mysql container. This
simplifies the deployment, as we no longer have an empty mysql-client
container sitting around.
When mysql is not enabled (ACTIVATE_MYSQL=False) the mysql container is
simply a mysql client.
