Deferred fetching

[noamr]

Add a JS-exposed function to request a deferred fetch, called fetchLater.

A deferred fetch would be invoked in one of two scenarios:

• The document is destroyed (the fetch group is terminated)
• A given period of time has passed.

A few constraints:

• Request body streams are not allowed
• This is only allowed in documents, and only reporting to potentially-trustworthy URLs
• Deferred fetch requests are limited to 64KB per origin. Exceeding this would immediately throw.

The quota algorithm is a bit intricate, but its default should be somewhat reasonable for all but advanced cases.

• A top level document has 640kb of quota for deferred fetching. This is important to avoid wasting high bandwidth after a tab has been closed. This quota is shared, by default, with the top-level's document same-origin same-agent descendants. The same-agent restriction is important for avoiding race conditions, as same-agent frames are guaranteed to call fetchLater in sequence.
• By default, 128kb out of the 640kb quota is reserved for cross-origin or cross-agent iframes. Permissions policy (deferred-fetch-minimal) controls that, and the top-level document can disable that allocated quota by disabling that permissions policy.
• Any document can delegate 64kb out of its reserved quota for cross-origin or cross-agent subframes, by explicitly enabling the deferred-fetch permissions policy.
• Reserving some of the quota to a cross-origin or cross-agent subframe happens when the frame is being navigated by the container, e.g. setting src on an iframe. It is not guaranteed that the subframe would actually be able to use that
  quota, as it might end up navigating to a same-origin URL or disable the feature in its own permissions policy. However, the container's document only cares about the initial reserved value for subframes it doesn't have direct access to.

See WICG/pending-beacon#70

• At least two implementers are interested (and none opposed):
  • Chromium (implementing)
  • WebKit (supportive, see fetchLater() API (previously: PendingBeacon API) WebKit/standards-positions#85)
  • Mozilla (supportive, see fetchLater() API mozilla/standards-positions#703)
• Tests are written and can be reviewed and commented upon at:
  • https://wpt.fyi/results/fetch/fetch-later?label=experimental&label=master&aligned
• Implementation bugs are filed:
  • Chromium: https://issues.chromium.org/issues/40276121
  • Gecko: https://bugzilla.mozilla.org/show_bug.cgi?id=1936180
  • WebKit: https://bugs.webkit.org/show_bug.cgi?id=284347
• MDN issue is filed: This is automatically created by the Chromium process when close to shipping.

(See WHATWG Working Mode: Changes for more details.)

---

Preview | Diff

[annevk]

1. "inactive timeout" -> "deferred delay". At least that seems clearer to me.
2. I think fetchLater was quite nice in that it sorts the same as fetch. fetchDeferred could also work, though is a bit harder to spell. I don't think we need request in the name.
3. It's not clear how the name fetch group states get activated. That needs some kind of additional PR against HTML I suppose?
4. I think we should describe the API in the same section as the fetch method. Could be called "Fetch methods" then.
5. Deferred fetching itself could then precede the "Fetch API" section. Maybe it could even be a subsection of "Fetching" though I don't mind a new top-level section.

[noamr]

1. "inactive timeout" -> "deferred delay". At least that seems clearer to me.

But this is deferred delay specifically in the case of inactivity. "inactivity deferred delay"?

2. I think fetchLater was quite nice in that it sorts the same as fetch. fetchDeferred could also work, though is a bit harder to spell. I don't think we need request in the name.

I was thinking about "what are we doing right now?" which is requesting/scheduling a deferred fetch for later. But fetchLater is also fine with me, it's very easy to remember.

3. It's not clear how the name fetch group states get activated. That needs some kind of additional PR against HTML I suppose?

Yes, HTML would activate/deactivate in the BFCache code path. Need to prepare a PR for that but wanted to see that I'm on the right track first.

4. I think we should describe the API in the same section as the fetch method. Could be called "Fetch methods" then.

Will do

5. Deferred fetching itself could then precede the "Fetch API" section. Maybe it could even be a subsection of "Fetching" though I don't mind a new top-level section.

OK

[mingyc]

cc @mingyc @fergald @yoavweiss @clelland latest API shape proposal for PendingBeacon

[mingyc]

Deferred fetch body sizes are limited to 64KB per origin. Exceeding this would immediately reject with a QuotaExceeded.

Another note about "origin" of a beacon request: there were some previous discussion about using 3P storage partitioning key (not origin, which is stricter) to decide whether pending beacon requests in a page are sendable or not in terms of privacy concern, see WICG/pending-beacon#30 (comment) and comments there below. I am not sure how this should be spec.

[noamr]

Deferred fetch body sizes are limited to 64KB per origin. Exceeding this would immediately reject with a QuotaExceeded.

Another note about "origin" of a beacon request: there were some previous discussion about using 3P storage partitioning key (not origin, which is stricter) to decide whether pending beacon requests in a page are sendable or not in terms of privacy concern, see WICG/pending-beacon#30 (comment) and comments there below. I am not sure how this should be spec.

OK, perhaps the 64kb constraint can be per network partition key rather than origin.

[mingyc]

PTAL, added some more comments.

[mingyc]

noamr@ PTAL. I've added some more questions and comments. Really thanks for your help!

[mingyc]

Thank you noamr@! I've added some last comments. And also added the missing features from the original proposal belows. Please let us know if they are suitable here.

1. It is suggested to allow only secure requests for new API in Do we want to enforce HTTPS request? WICG/pending-beacon#27. Should we enforce HTTPS-only requests on fetchLater()?
2. Should this spec mention [Permission Policy]https://www.w3.org/TR/permissions-policy/? In [Fetch-Based API] Permissions Policy WICG/pending-beacon#77, the suggestion is to allow the API by default. But we might want to provide a way to manage 3rd party iframe's usage.
3. Consider to support retry mechanism WICG/pending-beacon#40 Should this spec mention retry when fetchLater() fails to send/commit?
4. The original PendingBeacon proposal also includes Crash recovery WICG/pending-beacon#34, not sure how it can be incorporated into fetch spec.

[noamr]

Thank you noamr@! I've added some last comments. And also added the missing features from the original proposal belows. Please let us know if they are suitable here.

1. It is suggested to allow only secure requests for new API in Do we want to enforce HTTPS request? WICG/pending-beacon#27. Should we enforce HTTPS-only requests on fetchLater()?

Probably a good idea, from the point of view of enabling new features only for secure requests.

2. Should this spec mention [Permission Policy]https://www.w3.org/TR/permissions-policy/? In [Fetch-Based API] Permissions Policy WICG/pending-beacon#77, the suggestion is to allow the API by default. But we might want to provide a way to manage 3rd party iframe's usage.

I don't think we should integrate with permission policy. But we should allow the user agent to deny a fetchLater and throw an error immediately. I'll add that to the PR.

3. Consider to support retry mechanism WICG/pending-beacon#40 Should this spec mention retry when fetchLater() fails to send/commit?

Perhaps consider adding this later?

4. The original PendingBeacon proposal also includes Crash recovery WICG/pending-beacon#34, not sure how it can be incorporated into fetch spec.

I don't think that changes anything in the spec.

[mingyc]

Deferred fetch body sizes are limited to 64KB per origin. Exceeding this would immediately reject with a QuotaExceeded.

Another note about "origin" of a beacon request: there were some previous discussion about using 3P storage partitioning key (not origin, which is stricter) to decide whether pending beacon requests in a page are sendable or not in terms of privacy concern, see WICG/pending-beacon#30 (comment) and comments there below. I am not sure how this should be spec.

OK, perhaps the 64kb constraint can be per network partition key rather than origin.

@noamr Following up on the sendable beacon discussion:

As mentioned in WICG/pending-beacon#30 (comment), there were discussions around whether a beacon (or deferred request) should be sent when network changes. I tried to summarize them in [this PR](WICG/pending-beacon@feb3cf9, but basically to process a beacon request when BackgroundSync is off, we need to see if another open document (tab/frame/etc) with the same storage partitioning key as the current document's one, to avoid unexpected sending the request after network changes.

Do you think the above makes sense to be integrated into Fetch spec?

[mingyc]

This is called when container and the document that initiated the navigation (the "source document") are same origin.

It potentially reserves either 64kb or 8kb of quota for the frame, if it is not same origin with its parent and the permissions policy allow.

Could you please verify the triggering condition? The first sentence mentions same-origin only, while the second sentence still brings up if it is not same origin with its parent.

Even if the "not same origin" condition only applies to the 8kb case, when should it be triggered? Is it when different-origin container gets created or starts to navigate? But a different-origin container will get its initial policy copied from parent container, which it should not modifiy?

Given the following example

<!-- a.com/index.html -->

let iframe = document.createElement("iframe");
iframe.src = "b.com";
document.body.appendChild(iframe)

<!-- b.com/index.html -->
<script>
fetchLater("b.com");
</script>

• When iframe is created, it does not have any parent container.
• When iframe.src is set to b.com, it navigates to a cross-origin destination, with FramePolicy copied from parent document (?)
• When should iframe triggers the reserve deferred policy quota algorithm?

[mingyc]

If quota is less than quotaForOrigin, then return quota.

(Discussed offline in chat, but copying here)

Given the following examples:

// root (a.com) -> frame-1 (a.com) -> frame-2 (b.com)
//              -> frame-3 (b.com) -> frame-4 (a.com)

I'd like to verify quotas for the above 5 documents:

root: 64kb? (the result I got, but isn't 8kb taken by frame-2?)
frame-1: 54kb
frame-2: 8kb
frame-3: 8kb
frame-4: 0kb?

The questionable part is root and frame-1, they can't get more than 64kb quota due to the quoted return step at the beginning.