Requests for __data.json should be skipped entirely when possible

[Conduitry]

Describe the bug

Re-running server-side load on each navigation is supposed to be skipped now (unless the load function depends on url etc.), but this does not appear to be fully working. Even if a given page has no +page.server.js file, the browser is still making requests to __data.json on each navigation, which return things like {"type":"data","nodes":[{"type":"skip"},null]} - the root (layout) node is skipped, and the leaf (page) node has no data. Requests that would tell the server to skip all of the nodes for which there are no .server.js files should be skipped entirely.

Reproduction

https://github.com/Conduitry-Repros/kit-6201

Logs

No response

System Info

System:
    OS: Linux 5.10 Ubuntu 20.04.4 LTS (Focal Fossa)
    CPU: (8) x64 Intel(R) Core(TM) i7-10510U CPU @ 1.80GHz
    Memory: 8.34 GB / 12.32 GB
    Container: Yes
    Shell: 5.0.17 - /bin/bash
  Binaries:
    Yarn: 1.22.19
    npm: 8.15.0
  npmPackages:
    @sveltejs/adapter-auto: next => 1.0.0-next.65 
    @sveltejs/kit: next => 1.0.0-next.431 
    svelte: ^3.44.0 => 3.49.0 
    vite: ^3.0.4 => 3.0.9

Severity

serious, but I can work around it

Additional Information

This is the situation I referred to in #6056 (comment) - it seems like it was not correctly (or fully?) handled.

[Conduitry]

An intriguing wrinkle with this is that, by calling __data.json on every navigation whether it needs to be or not, any logging of requests in handle will also reflect any client-side navigation through the app, and (since /__data.json is stripped off from the URL in the event), these logs will look pretty much exactly the same as if each request were being server-rendered. This might be good or it might be bad.

[bjon]

@Conduitry __data.json been stripped off is really bad for us. We have to use event.request.url, cause there you can see it.

[dummdidumm]

What's your use case for knowing that it's the __data.json call?

[stalkerg]

yep, extra __data.json loaded even for preloading, which has no sense to me. In my case, I have 300ms network latency and an extra request is very sensitive.

[bjon]

@dummdidumm We don't want to know. Its the effect of removing __data.json, that's the problem.

Here's how our URLs works (predated sveltekit):
/[language]/ is our "root". /en/path1 and /en/path2/path3 are valid, cause we only use trailing slash for language path.
If we detect /en, we redirect to /en/ , and this is the problem.
When /en/__data.json is requested, event.url becomes /en, and we do a redirect, that breaks the app.
We have solved this by never using event.url, but using event.request.url instead.

Is there a good reason why event.request.url and event.url should differ? To me is seem strange that we can not "trust" event.url, to be the requested url.

[dummdidumm]

One problem with calling the backend too much is that we don't know that the new page doesn't have a load function. We only know "this route uses server data somewhere along the way", but we don't know "this leaf doesn't use a load function and everything leading up to it is unchanged so you can skip requesting the data".

[stalkerg]

@Conduitry @dummdidumm after the latest changes, this issue occurs again.