Html: Script modules and credentials

[jakearchibald]

By default <script type="module"> will omit credentials for all module fetches, much like the default for fetch().

However…

<!-- No credentials -->
<script type="module" src="foo.js"></script>
<!-- With credentials -->
<script type="module" src="foo.js" crossorigin></script>

It feels really weird that crossorigin implies adding credentials for same-origin requests.

[domenic]

This may be clearer if you stop relying on the invalid value default and type crossorigin="anonymous".

We've talked about introducing a new attribute, credentials=omit|same-origin|include (with invalid/missing value default of omit), but I'm not sure it pulls its own weight.

[annevk]

My tentative suggestion is that we try to flip the default of fetch() and this to "same-origin" and keep "omit" only in fetch() for the extra-paranoid. That simplifies our default setups a bit while still allowing for all the things (albeit some only through service workers). My hope is that the compatibility impact of this will be low, since adding credentials is less problematic than removing them, and that the ensuing developer experience will be much better, since they have one less default to learn.

[domenic]

One concrete downside there is that it means the many requests per day made with fetch() and modules will now have the extra cookie bytes sent along, which can be quite significant in some cases. But I guess maybe that's minor in compared to the confusion here? I dunno, I'd say introduce a new content attribute if it's really that bad.

[jakearchibald]

@domenic

This may be clearer if you stop relying on the invalid value default and type crossorigin="anonymous".

It still feels weird. Without crossorigin, it's anonymous, and if you want to it be not-anonymous to the same origin to tell it to be anonymous cross origin. I guess it's proving the rule with an exception, but it feels a bit backwards.

One concrete downside there is that it means the many requests per day made with fetch() and modules will now have the extra cookie bytes sent along

On the other hand, those requests will be able to use the HTTP/2 connection already opened by the page request, whereas non-credentialed requests require another connection.

Changing fetch's default feels like a huge change, but it does catch people out all the time.

[jakearchibald]

This impacts HTTP/2 push as well. If you push your JS along with the page, it'll only be used if the developer opts into credentials for their module scripts.

I was debugging a case recently where a developer was pushing resources along with their page (or service worker, I can't remember) and wondering why they weren't being picked up when they did cache.addAll().

[annevk]

I think the new HTTP connection coupled with expectations about the default is more severe than a couple of bytes that could end up compressed in H/2 if they are the same for each request (as a single-byte pointer per https://tools.ietf.org/html/rfc7541 if I understand things correctly).

[domenic]

Just so people are aware, my position is that I would like module scripts to follow fetch(). If fetch changes its defaults, then I think module scripts should too. But if fetch() does not, I'd like to stick with what we have, even with the weird naming.

Is anyone interested in driving a decision on the fetch side?

[annevk]

Created https://github.com/whatwg/fetch/pull/585.

[annevk]

Fetch is updated now and implementations are expected to align. We should update module scripts too.

[annevk]

I looked at this and I think we should tackle this together with module workers and no longer allowing CORS for them. Then we can also remove the credentials member there entirely.

[annevk]

I didn't find time to untangle the entire script loading setup to make the changes needed here. https://github.com/whatwg/html/tree/annevk/module-scripts-and-credentials has some WIP.

[annevk]

Hmm, the top-level fetch needs to be same-origin, but what about the imports? importScripts() can go cross-origin, so I guess they should be able to do that too? Basically the same as a service worker?

It's probably best if @domenic fixes this since he has a much better handle on the various algorithms, but I'm willing to do the work given some feedback on all of the above.

[domenic]

Help fixing this would be much appreciated.

I thought the idea was to be consistent with fetch()'s default. In which case, it should be "same-origin" always?

Or are you asking about allowing them to have the option of "include"? For that case, yeah, I'd suggest keeping the credentials member for WorkerOptions, instead of removing it, and just changing the default.

[domenic]

Note also that importScripts() does not work inside a module worker. (But import does.)

[annevk]

Basically, I'm wondering if we should fix #3109 as part of this, since it makes credentials mode less interesting for module workers, at least for the top-level fetch. It's a little bit unclear if the imports for such a module worker can still go cross-origin though; I guess in theory there might not be a problem with that.

If we don't couple them, then I guess we need to keep the credentials member, but we should not allow it to be set to "omit" as you cannot get that for module scripts either.