🦋 Changeset detected

Latest commit: c1ea90a

The changes in this PR will be included in the next version bump.

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

Thanks for finding and fixing this!

It looks to me like this happens on the client-side JS. I can't reproduce it on the SvelteKit site and I'm guessing that may be because we only do manual imports there. I'm really curious why it doesn't happen with manual imports, but does happen when simply sticking the path in the src. This PR has a small performance hit, so we should investigate a bit to see if we can avoid the import.meta.url stuff from being generated in the first place.

Here's the relevant location in Vite:

https://github.com/vitejs/vite/blob/b93dfe3e08f56cafe2e549efd80285a12a3dc2f0/packages/vite/src/node/plugins/asset.ts#L132

Hmm. In #9220 we seem to be forcing relative paths (which was attempted to be changed and then later reverted in #10442). I don't like the API of paths.relative: undefined | false, which I find to be a bit confusing. I don't understand exactly what the undefined case is trying to solve for. It sounds like it might help the Internet Archive? I'd expect the Internet Archive to be able to deal with all the sites out there that are not coded especially for it with a weird half-relative and half-absolute behavior, but maybe that's overly optimistic.

The undefined value causes hydration slowdowns due to different values on the client and server requiring complex comparisons, causes larger bundle sizes from inserting import.meta.url everywhere, and now is causing template creation to be de-optimized. I really dislike it. I'd love to just see a default value of false with undefined removed in Kit 2.0, but am not sure if that's feasible.

We will need a solution like in this PR regardless of whether we can keep Kit in an optimized state because you'll end up with this code when the user sets relative: true. But we should only use a mustache expression when dealing with relative paths. And I do also wonder if we can at least strip the leading "" + and trailing + "" to be slightly more optimized about it in that case.

I also really dislike this solution, but given that the alternative is all images being broken on pages with CSR, might it be better to merge this ASAP and then look into performance improvements?