fix: await requests to before server restart

[patak-dev]

Description

When a server is restarted, server.moduleGraph will end up being a new clean graph. But server a single object, that we keep across restarts. Outdated plugins and middlewares may be processing requests and if they use server.moduleGraph they could crash or add outdated modules or relationships to the new graph.

This PR awaits processing to be done before doing Object.assign(server, newServer)

We were already awaiting the pluginContainer to close, but this was only calling buildEnd and closeBundle. It nows await for all hooks to finish.

We also await from server._pendingRequests.

See #13231 (comment)

We also discussed with @antfu that a Proxy could be used to give the plugins a view that keeps server.moduleGraph pointing to the old graph. The server proxy that plugins will get in configureServer will be this proxy. I think this could be worth exploring, but if we properly await for the plugins to finish processing, we shouldn't need to do it.

What we may be missing now is to await for external middlewares. We can check this in a future PR.

---

What is the purpose of this pull request?

• Bug fix
• New Feature
• Documentation update
• Other

[stackblitz]

Run & review this pull request in StackBlitz Codeflow.

[patak-dev]

Fixed now: I don't know yet why this is failing, I'll review it later. If you spot the issue, let me know :)

[patak-dev]

/ecosystem-ci run

[vite-ecosystem-ci]

📝 Ran ecosystem CI: Open

suite	result
astro	✅ success
histoire	✅ success
iles	✅ success
ladle	✅ success
laravel	✅ success
marko	✅ success
nuxt	❌ failure
nx	✅ success
previewjs	✅ success
qwik	✅ success
rakkas	✅ success
sveltekit	✅ success
unocss	✅ success
vite-plugin-ssr	✅ success
vite-plugin-react	✅ success
vite-plugin-react-pages	✅ success
vite-plugin-react-swc	✅ success
vite-plugin-svelte	✅ success
vite-plugin-vue	✅ success
vite-setup-catalogue	✅ success
vitepress	✅ success
vitest	✅ success

[patak-dev]

We could revert #13231 after this PR, as old plugins should always see the correct server.moduleGraph but it is probably ok to leave the guard. Maybe we should merge this PR in the beta for 4.4 just in case the changes in server.restart() affect some downstream deps we aren't testing in ecosystem CI.

Note: Nuxt is expected to fail, it is also failing in main now.

[benmccann]

I wonder if it's worth having a method like server.throwIfClosed() to encapsulate this logic and potentially avoid needing to access a private variable. not sure if you'd want to expose this externally, but if so then middlewares could call it if necessary

[patak-dev]

That's an interesting idea. I think we better explore it in a separate PR, and once we have a use case from a middleware to justify exposing the function. transformRequest is internal to vite so I think it is fine reading _restartPromise here for now, and we may find a better way to stop the execution later. Throwing here feels a bit hacky to me to expose it.

[benmccann]

I left a few small comments, but looks good to me!

[bluwy]

Some tiny nits. Looking forward to trying this in the minor!

[bluwy]

Maybe we should also rebase the branch before merging just in case we miss a change from main.