> This probably requires an Enterprise Policy, to reduce the risk for
managed installs. +bheenan@ for opinions on that front.

I agree, this looks like a breaking change according to
go/chrome-enterprise-friendly and therefore needs a policy. Instructions
for implementing a policy are here:
https://source.chromium.org/chromium/chromium/src/+/main:docs/enterprise/add_new_policy.md
if you haven't done it before, and the enterprise team is happy to help if
anything seems confusing. Having this implemented as a "soft removal" with
a temporary policy escape hatch significantly reduces enterprise risk,
since even if we are surprised by a use case hiding behind many
enterprises' tendency to turn off metrics, those users can
break-fix themselves immediately while staying on the latest version.

On Wed, Jan 12, 2022 at 12:45 PM Yoav Weiss <[email protected]> wrote:

> Hey Daniel!
>
> While searching for this intent review, I stumbled upon
> https://developer.chrome.com/blog/immutable-document-domain/
> That's a useful piece of documentation! Thanks +Eiji Kitamura!!
>
> This intent was just discussed at the API owner meeting (where Chris,
> Rego, Daniel, Philip, Alex, MikeT and myself were present).
> This change seems risky in terms of potential breakage when looking at our
> stats, and that's even before talking about enterprises, where a lot of the
> API owners feel the risk is even higher.
>
> Given that, here's a few potential next steps to try and reduce that risk:
>
>    - UKM and outreach to specific large users of the API can maybe help
>    drive the usage down.
>    - A deprecation period of 3 milestones feels a bit short here. Is the
>    expectation that turning on the opt-out header can be done under that
>    period?
>    - A report-only mode could have allowed sites to try and enable this,
>    without risking actual breakage for their documents/properties that use
>    document.domain. This is doubly true for platforms that want to warn their
>    customers about this upcoming deprecation, but without taking risks on
>    their behalf. At the same time, it is true that they could collect
>    deprecation reports (thanks for adding those!) instead during the
>    deprecation period, which can be considered an on-by-default report-only
>    mode. Can y'all add specific guidance on deprecation reports to the
>    documentation?
>    -  It'd be helpful to reach out to enterprise folks and see what their
>    responses may be for this. +Greg Whitworth.
>    - This probably requires an Enterprise Policy, to reduce the risk for
>    managed installs. +bheenan@ for opinions on that front.
>    - Is there a plan to eventually remove the opt-out option? Or is it
>    the plan to have it in place permanently?
>
> Cheers,
> Yoav
>
>
> On Saturday, December 18, 2021 at 9:38:33 PM UTC+1 Mike Taylor wrote:
>
>> On 12/16/21 5:52 PM, Charlie Reis wrote:
>>
>>
>>
>> On Thu, Dec 16, 2021 at 7:28 AM 'Daniel Vogelheim' via blink-dev <
>> [email protected]> wrote:
>>
>>> On Tue, Dec 14, 2021 at 11:51 PM Mike Taylor <[email protected]>
>>> wrote:
>>>
>>>> On 12/14/21 11:35 AM, Daniel Bratell wrote:
>>>>
>>>> It seems more or less everyone agrees on this being a good thing, so it
>>>> mainly comes down to web compatibility.
>>>>
>>>> How much of the web will break, and how badly. The numbers mentioned,
>>>> 0.5% of sites set document.domain, 0.05% seem to depend on document.domain,
>>>> are quite high. One thing I didn't quite pick up is what happens if you try
>>>> to set document.domain when it's not settable. Will it silently pretend to
>>>> work, or will that also be a possible break point?
>>>>
>>>> I would be surprised if it didn't behave the same as setting an
>>>> arbitrary expando on `document` - we should be safe there.
>>>>
>>>
>>> Almost. The error handling is mostly the same. But while a foreign
>>> attribute on document would return the new value, document.domain (when in
>>> an origin-keyed agent cluster) does not.
>>>
>>> So:
>>>   document.foo = "bla"
>>>   document.foo  // Returns "bla".
>>>
>>>   document.domain = "bla"  // Throws SecurityException.
>>>
>>>   // On a domain www.host.com, site-keyed agent cluster (current
>>> default)
>>>   document.domain = "host.com"  // Succeeds.
>>>   document.domain;  // returns "host.com".
>>>
>>>   // On a domain www.host.com, origin-keyed agent cluster (future
>>> default)
>>>   document.domain = "host.com"  // Doesn't throw. Doesn't do anything
>>> else either.
>>>   document.domain;  // Still returns www.host.com.
>>>
>>> Risks: Interoperability and Compatibility
>>>>
>>>> * No interoperability risks. *
>>>> Compatibility risk exists, but is fairly small as document.domain is an
>>>> already deprecated feature. We’ve detailed UKM metrics in place and are
>>>> planning to reach out to top users as soon as we’ve LGTMs for the plan.
>>>>
>>>> As Daniel mentioned, I think the compat risk should be considered to be
>>>> higher, despite this being deprecated for a long time.
>>>>
>>>
>>> Yes, agreed.
>>>
>>>> Current usage of the document.domain: 0.05%
>>>> <https://chromestatus.com/metrics/feature/timeline/popularity/2544> of
>>>> page views rely upon document.domain to enable some cross-origin access
>>>> that wasn't otherwise possible. 0.24%
>>>> <https://chromestatus.com/metrics/feature/timeline/popularity/2543> of
>>>> page views block same-origin access because only one side sets
>>>> document.domain. Both counters can be found on
>>>> https://deprecate.it/#document-domain, too.
>>>>
>>>> (cool site, btw)
>>>>
>>>> We’ve reached out in advance to 4 of the top current users - TL;DR Most
>>>> of their use cases could be achieved already by different APIs e.g.
>>>> Audio/video autoplay in iframes by adding the ‘autoplay’ attribute, support
>>>> chat deployed across subdomains.
>>>>
>>>> Out of curiosity, did any of them mention what couldn't be achieved via
>>>> existing APIs?
>>>>
>>>
>>> I checked back, and nothing particular came up. It seems that migrating
>>> off of document.domain isn't blocked by a lack of options.
>>>
>>> Activation - Deprecation plan
>>>> M98 - Add the devtools issue and warning incl. a web.dev blog post to
>>>> guide adoption
>>>>
>>>> * M98-M100 - Monitor use counters and reach out to current users *
>>>>
>>>> What's the plan if the use counters don't move? Do you have a minimum
>>>> page view % in mind you would want before proceeding to the next step (even
>>>> if it meant delaying the timeline)?
>>>>
>>>
>>> We don't have a dead-set plan. The primary idea is a combination of
>>> delay-ing until usage is low enough, and outreach to offending sites to
>>> educate about the problem & available alternatives.
>>>
>>>
>>>
>>>> * M101 - Deprecate the feature by default. No reverse-origin trial is
>>>> planned as the ‘Origin-Agent-Cluster’ http header can be used to gain
>>>> access to the feature. *
>>>>
>>>> Would this disabled-by-default change ride the trains, or have you
>>>> considered finching it out to assess compat risk?
>>>>
>>>
>>> My original plan was to enable it by default in the 101 release, and
>>> have a Finch switch as an emergency-off. Reading the feedback here, maybe
>>> it's better to incrementally enable it via Finch. I'll be happy to follow
>>> whatever path API owners prefer.
>>>
>>
>> In my experience (caveat: I'm not an API owner), it's harder to repro and
>> triage compatibility bugs that get filed if a feature is only enabled for a
>> percentage of users via Finch.  It tends to be quicker to bisect reports
>> and get attention on a compat bug if the feature is enabled on trunk
>> instead, with an easy way to revert if needed.  (Finch is certainly better
>> when you care about performance issues, which doesn't seem to be the case
>> here.)
>>
>> Yeah, I hear you - the unpredictability is a challenge. My preferred
>> approach would be to hold things at 100% in pre-release channels for some
>> period of time to sniff out compat bugs - but AIUI, this isn't really a
>> thing that Finch is designed to do, and pre-release comes with its own set
>> of biases that may not accurately reflect release behavior. But where the
>> risk is non-zero, only breaking some users seems better than breaking all
>> users, even if imperfect.
>>
>

-- 
You received this message because you are subscribed to the Google Groups 
"blink-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
To view this discussion on the web visit 
https://groups.google.com/a/chromium.org/d/msgid/blink-dev/CAFKd2qXjZWrGgCAPzV0ck74SLPM5vLNb0y-mx772OdQ%2B-zGEaA%40mail.gmail.com.

Reply via email to