OT Limit: Speculation Rules/Prefetch

34 views
Skip to first unread message

Jeremy Roman

unread,
Feb 25, 2022, 4:21:14 PM2/25/22
to blink-api-owners-discuss, Michael Buettner, Kenji Baheux
Hello API owners,

The speculation rules experiment has been flagged as exceeding the usage limit, presumably due to the private prefetch proxy Early Access Program which uses the feature.

I don't believe this risks burn-in and we should not disable the trial, for the following reasons:

1. The effective usage is likely much lower than the UseCounter suggests. At the moment only certain origins that have opted in can receive prefetch proxy traffic, but this detection happens after the UseCounter for using speculation rules has fired.

Ideally we would have a UseCounter for not only the use of the speculation rules syntax, but for that syntax actually resulting in a connection to the origin server. We can look into that if API owners want (it shouldn't be terribly complicated) but it's not something that exists today.

2. This usage arises from a partner with whom we are coordinating closely and can modify or revert their usage of the feature. This tends to reduce the risk of burn-in, as opposed to usage by a large number of developers who could not make such changes.

3. Unshipping the feature would not break sites. Since this prefetching is a performance enhancement and calling code has no effect in browsers without this shipped, unshipping it even if developers don't remove their usage won't break the web.

4. Disabling the trial would reduce our ability for developers to gather data about the impact of this feature. We'd prefer not to be disruptive if possible.

For these reasons, I believe the OT usage limit to prevent burn-in is inapplicable here and request an exception.

Chris Harrelson

unread,
Feb 25, 2022, 4:26:55 PM2/25/22
to Jeremy Roman, blink-api-owners-discuss, Michael Buettner, Kenji Baheux
Hi Jeremy,

Thanks for this info. Can you tell us what percentage of traffic you're requesting approval for?

--
You received this message because you are subscribed to the Google Groups "blink-api-owners-discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to blink-api-owners-d...@chromium.org.
To view this discussion on the web visit https://groups.google.com/a/chromium.org/d/msgid/blink-api-owners-discuss/CACuR13cyNfrytcq5s5KQunyKthept%3DLU%3DZ7o-AVbN3h0KdfxLQ%40mail.gmail.com.

Jeremy Roman

unread,
Feb 25, 2022, 6:37:53 PM2/25/22
to Chris Harrelson, blink-api-owners-discuss, Michael Buettner, Kenji Baheux
On Fri, Feb 25, 2022 at 4:26 PM Chris Harrelson <chri...@chromium.org> wrote:
Hi Jeremy,

Thanks for this info. Can you tell us what percentage of traffic you're requesting approval for?

As a percentage of pages containing speculation rules over page loads overall, on all platforms?

I think 5% would suffice, bearing in mind that the significant majority of those page loads don't actually result in a connection to an origin server (based on back-of-the-envelope calculations in consultation with Simon, using the metrics we do have).

Jeremy Roman

unread,
Mar 2, 2022, 10:23:05 AM3/2/22
to Chris Harrelson, blink-api-owners-discuss, Michael Buettner, Kenji Baheux
I realize it's a busy time, but when can we expect to hear back on this?

Jeremy Roman

unread,
Mar 2, 2022, 5:00:24 PM3/2/22
to Chris Harrelson, blink-api-owners-discuss, Michael Buettner, Kenji Baheux
To elaborate, under the assumption this partner constitutes the vast majority of usage, about 1% of current usage from the main caller is the pre-existing experiment and the remaining 99% is the private prefetch proxy EAP (destinations must opt in to participate in the early access program). The 1% experiment dominates prefetches which are actually executed, on about 0.02% of page loads (estimated from the total usage at issue, since we don't directly measure this) at present. If more publishers joined the EAP that number could rise, but I wouldn't expect it to reach 0.5% during the experiment period.

Chris Harrelson

unread,
Mar 2, 2022, 5:26:37 PM3/2/22
to Jeremy Roman, blink-api-owners-discuss, Michael Buettner, Kenji Baheux
I think what you're saying is that while about 5% of total page loads in Chrome would be potentially using speculation rules in the sense of being allowed to use it, but in practice only about 0.02% would currently be doing so, and even if a lot more publishers joined you can't see it going about 0.5%. 

From my perspective this sounds fine, and below the threshold for regular OTs because of this analysis. Any other opinions?

Jeremy Roman

unread,
Mar 2, 2022, 5:28:13 PM3/2/22
to Chris Harrelson, blink-api-owners-discuss, Michael Buettner, Kenji Baheux
On Wed, Mar 2, 2022 at 5:26 PM Chris Harrelson <chri...@chromium.org> wrote:
I think what you're saying is that while about 5% of total page loads in Chrome would be potentially using speculation rules in the sense of being allowed to use it, but in practice only about 0.02% would currently be doing so, and even if a lot more publishers joined you can't see it going about 0.5%. 

Yes, that's a good summary.

Yoav Weiss

unread,
Mar 3, 2022, 1:03:35 AM3/3/22
to Jeremy Roman, Chris Harrelson, blink-api-owners-discuss, Michael Buettner, Kenji Baheux

Chris Harrelson

unread,
Mar 4, 2022, 4:46:02 PM3/4/22
to Yoav Weiss, Jeremy Roman, blink-api-owners-discuss, Michael Buettner, Kenji Baheux
Hearing no other objection, consider your OT approved to go up to 5% "nominal" coverage, but only so long as your statistics indicate the *actual* usage is below 0.5%. If it ever goes above that, we'll need 3 LGTMs.

Reply all
Reply to author
Forward
0 new messages