Ecto Sandbox ownership issue when persisting state in terminate callback

Hello,

Context
I am developing a toy website to get myself familiar with Elixir and Phoenix. The website is like multiple shared canvases where everyone with a token can change pixels (a bit like multiple tiny https://pixelcanvas.io/). On the home page, you can see auto-refreshing thumbnails of each canvas, and when you click on one, you see the full-size version with live updates using a Phoenix Channel. The canvases can be created or deleted on the fly in the admin section, which will create the new canvas (UUID, name) in the DB and spin-up a (unique) GenServer (per canvas). Without persistence, I successfully deployed it in a cluster, using global processes (via Horde), so far so good.

Now, I want to persist the images. To do that, whenever a canvas GenServer terminates, it saves the PNG blob in the DB. When it starts, it tries to load it from the DB. This works well, I can add/remove nodes from the cluster and the image survives.

Here is how canvas GenServers are handled:

• when I create a canvas using Canvases.create_canvas/1, the function creates the row in DB, starts the GenServer under a (Horde) DynamicSupervisor using CanvasManager.start_canvas_server/1, and publishes a :canvas_created PubSub event to update the LiveViews.
• when I delete a canvas, using Canvas.delete_canvas/1, it deletes the canvas and persisted images from the DB and publishes a :canvas_deleted PubSub event for both the LiveView and the CanvasManager, which will in turn terminate the GenServer.
• when the application starts, CanvasManager is started under the supervision tree and will start all canvases.

The GenServer is needed to interact with the canvas: get the version, get the image as PNG, get a pixel color, set a pixel color. The DB is only used to list the existing canvases and persist the image.

Problem
I have a hard time figuring out how to test the GenServer now that there is persistence. I understand the issue, but I don’t know what’s the best way to address it: when I create a new canvas, I call Canvases.create_canvas/1 via a fixture; this will also start the GenServer under my application’s supervisor (not ExUnit’s). To stop the GenServer a the end of the test, the fixture calls ExUnit.Callbacks.on_exit/1. The GenServer cannot work without the row in DB (will fail on termination), and I want to keep the supervision tree clean after each test. The issue is that on_exit is executed in a different process than the test one, which is not the owner of the sandbox DB connection.

I read in some issue on GitHub that Jose does not recommend updating the DB from inside on_exit, so I think that there is something wrong with my design. What would you suggest then?

• Manually waiting in every test that creates a canvas for the GenServer to terminate?
• Not starting GenServer in tests that don’t need it? How? And that does not solve the issue for the tests that do need it.
• Start the processes using ExUnit.Callbacks.start_supervised/2? How can I do that? Would that solve the ownership issue?
• Something else? I don’t have much ideas.

Please tell me if you need any more details. The full source code is on GitHub (“canvas” => “lobby”/“image”, the GenServer is called ImageServer, I renamed here for clarity).

Thanks for the help!
Thomas

* For the curious: I am using a straightforward Rust NIF via Rustler to be able to mutate images.