feat!: don't run Celery workers in dev mode

[arbrandes]

Tutor's importing * from devstack.py[1] for the development settings, and that means that we aren't using Celery workers at all in dev mode (see [2]). This removes them from the dev compose file, thus saving everyboding a significant chunk of RAM.

[1] https://github.com/overhangio/tutor/blob/master/tutor/templates/apps/openedx/settings/lms/development.py#L3
[2] https://github.com/openedx/edx-platform/blob/master/lms/envs/devstack.py#L35

To do so, we rely on Docker compose profiles[3].

[3] https://docs.docker.com/compose/profiles/

BREAKING CHANGE: the COMPOSE_PROJECT_STARTED hook signature had to be changed to accomodate profile selection.

[DawoudSheraz]

We would need some sort of guidelines to understand how the workers can be setup in dev mode, especially now that we are using docker compose profiles.

[arbrandes]

Just a simple tutor dev dc start lms-worker would start the service, for instance, but the workers wouldn't be actually usable without modifying settings.py. Is this what you mean we should document?

Of course, the other question would be: why would anybody want to do this in a dev environment?

[DawoudSheraz]

Yeah, the command and setting changes, yes. I get the point that most devs might not be using this, we can have the doc in a followup. It is not a blocker.

[regisb]

I wasn't familiar with profiles, to be honest. I like the idea of not running workers in dev, but I'd like to avoid making a breaking change to the filter API, and also changes across many python functions.

Instead, could we simply move the lms-worker and cms-worker declarations from docker-compose.yml to docker-compose.prod.yml?

[arbrandes]

I don't think it's going to be that simple. There are other places in the code that assume those service definitions exist. For instance, https://github.com/overhangio/tutor/blob/master/tutor/commands/compose.py#L289. Plus, I suspect profiles might be useful down the road, woudn't you say so?

If you're dead set against it, though, there is another way: we can just set the worker services in the dev compose override file to a profile like "donotstart". We get the same thing this PR provides, but without actually supporting different profiles.

[DawoudSheraz]

The changes look good to me, though I will keep this open until Régis is back.

[kdmccormick]

Haha, I have a competing proposal (although I let it languish, because I was never able to confirm or deny that pdb would work on these workers):

• fix: run Celery tasks asynchronously in dev mode #928

My thinking was that it would be good to run these tasks asynchronously, because it would expose developers to the same race conditions that inevitably always happen with Celery tasks in production. But I can see the other side of the argument: async tasks are harder to debug and the workers use resources.

Either way, I think we can all agree that we should either use these workers, or turn them off. So I see four paths forward:

1. Use this PR, which will disable workers in dev. Open edX contributors who work on Celery tasks should take note to manually test their code using tutor local.
2. Use my PR, which will enable Celery tasks by default for everybody.
3. Create a new Tutor ASYNC_CELERY_TASKS configuration flag, enabled by default. Disabling it would remove the Celery workers from docker-compose and configure Django to run the tasks in-process.
4. Same as (3), but implement it as a tutor-celery plugin. Like tutor-mfe, it would be installed and enabled by default.

My vote is (1), but I would support any of them.

[DawoudSheraz]

Haha, I have a competing proposal (although I let it languish, because I was never able to confirm or deny that pdb would work on these workers):

• fix: run Celery tasks asynchronously in dev mode #928

My thinking was that it would be good to run these tasks asynchronously, because it would expose developers to the same race conditions that inevitably always happen with Celery tasks in production. But I can see the other side of the argument: async tasks are harder to debug and the workers use resources.

Either way, I think we can all agree that we should either use these workers, or turn them off. So I see four paths forward:

1. Use this PR, which will disable workers in dev. Open edX contributors who work on Celery tasks should take note to manually test their code using tutor local.
2. Use my PR, which will enable Celery tasks by default for everybody.
3. Create a new Tutor ASYNC_CELERY_TASKS configuration flag, enabled by default. Disabling it would remove the Celery workers from docker-compose and configure Django to run the tasks in-process.
4. Same as (3), but implement it as a tutor-celery plugin. Like tutor-mfe, it would be installed and enabled by default.

My vote is (1), but I would support any of them.
My take on all points:

1. This seems like the most viable solution. However, I also 2nd the fact that devs should be able to run the tasks async if they want to. If they can do so using local, good enough. But it will confuse devs as we suggest dev for primarily development purposes. Switching to local results in images being rebuilt and all, which can be infuriating if you are trying to debug.
2. Already discussed in 1
3. Régis has argued to avoid adding new configs and instead add filters/plugins if needed.
4. How do you envision the plugin would work/would be responsible for within the ecosystem? To relate with 1, if we can use the plugin for this purpose, it would be nice. I suspect we would be using the plugin to override IDA's celery settings and also allow celery configs to be added without modification to the core (sort-of similar situation Régis is describing on feat: Celery worker concurrency setting #1010 (comment))