RESOLVES FRMW-2978
**What**
Add retry mechanism to database connection management to prevent failing when the server start faster than what makes the connection available
**What**
Make sure there is no open handles left and that the shutdown function are properly called. Refactor and improve the medusa test runner. Make sure all modules instances are released and cleaned up
**NOTE:**
On a separate PR we can continue the investigation for the memory growing over time while the tests execute
**What**
- Reworks how admin extensions are loaded from plugins.
- Reworks how extensions are managed internally in the dashboard project.
**Why**
- Previously we loaded extensions from plugins the same way we do for extension found in a users application. This being scanning the source code for possible extensions in `.medusa/server/src/admin`, and including any extensions that were discovered in the final virtual modules.
- This was causing issues with how Vite optimizes dependencies, and would lead to CJS/ESM issues. Not sure of the exact cause of this, but the issue was pinpointed to Vite not being able to register correctly which dependencies to optimize when they were loaded through the virtual module from a plugin in `node_modules`.
**What changed**
- To circumvent the above issue we have changed to a different strategy for loading extensions from plugins. The changes are the following:
- We now build plugins slightly different, if a plugin has admin extensions we now build those to `.medusa/server/src/admin/index.mjs` and `.medusa/server/src/admin/index.js` for a ESM and CJS build.
- When determining how to load extensions from a source we follow these rules:
- If the source has a `medusa-plugin-options.json` or is the root application we determine that it is a `local` extension source, and load extensions as previously through a virtual module.
- If it has neither of the above, but has a `./admin` export in its package.json then we determine that it is a `package` extension, and we update the entry point for the dashboard to import the package and pass its extensions a long to the dashboard manager.
**Changes required by plugin authors**
- The change has no breaking changes, but requires plugin authors to update the `package.json` of their plugins to also include a `./admin` export. It should look like this:
```json
{
"name": "@medusajs/plugin",
"version": "0.0.1",
"description": "A starter for Medusa plugins.",
"author": "Medusa (https://medusajs.com)",
"license": "MIT",
"files": [
".medusa/server"
],
"exports": {
"./package.json": "./package.json",
"./workflows": "./.medusa/server/src/workflows/index.js",
"./.medusa/server/src/modules/*": "./.medusa/server/src/modules/*/index.js",
"./modules/*": "./.medusa/server/src/modules/*/index.js",
"./providers/*": "./.medusa/server/src/providers/*/index.js",
"./*": "./.medusa/server/src/*.js",
"./admin": {
"import": "./.medusa/server/src/admin/index.mjs",
"require": "./.medusa/server/src/admin/index.js",
"default": "./.medusa/server/src/admin/index.js"
}
},
}
```
Fixes: FRMW-2742
In this PR, we fix the build output of the backend source code, which eliminates a lot of magic between the development and production environments.
Right now, we only compile the source files from the `src` directory and write them within the `dist` directory.
**Here's how the `src` directory with a custom module looks like**
```
src
├── modules
│ └── hello
│ ├── index.ts
```
**Here's the build output**
```
dist
├── modules
│ └── hello
│ ├── index.js
```
Let's imagine a file at the root of your project (maybe the `medusa-config.js` file) that wants to import the `modules/hello/index` file. How can we ensure that the import will work in both the development and production environments?
If we write the import targeting the `src` directory, it will break in production because it should target the `dist` directory.
## Solution
The solution is to compile everything within the project and mimic the file structure in the build output, not just the `src` directory.
**Here's how the fixed output should look like**
```
dist
├── src
│ ├── modules
│ │ └── hello
│ │ ├── index.js
├── medusa-config.js
├── yarn.lock
├── package.json
```
If you notice carefully, we also have `medusa-config.js`, `yarn.lock`, and `package.json` within the `dist` directory. We do so to create a standalone built application, something you can copy/paste to your server and run without relying on the original source code.
- This results in small containers since you are not copying unnecessary files.
- Clear distinction between the development and the production code. If you want to run the production server, then `cd` into the `dist` directory and run it from there.
## Changes in the PR
- Breaking: Remove the `dist` and `build` folders. Instead, write them production artefacts within the `.medusa` directory as `.medusa/admin` and `.medusa/server`.
- Breaking: Change the output of the `.medusa/server` folder to mimic the root project structure.
- Refactor: Remove `Symbol.for("ts-node.register.instance")]` check to find from where to load the source code.
- Refactor: Use `tsc` for creating the production build. This ensures we respect `tsconfig` settings when creating the build and also perform type-checking.
Co-authored-by: Adrien de Peretti <25098370+adrien2p@users.noreply.github.com>
new wrapper for medusa integration tests.
for now it is only applied to the modules directory, but it could be used in the api integration tests or any other integrations that requires a db and a server up and running.
It is not perfect, but I wanted to have something working and centralised before improving it, also avoiding too many conflicts with other prs
