Re: Reprioritization - implementation intent

s/QUIC Working Group/HTTP Working Group/

On Tue, Jul 28, 2020 at 8:42 AM Yoav Weiss <yoav@yoav.ws> wrote:

> I ran the above by +David Schinazi <dschinazi@google.com>, who better
> represents the Chrome networking team than me. His reply is: "Chrome will
> closely follow proposals related to HTTP/3 reprioritization, and is very
> likely to implement a proposal that gets consensus in the IETF QUIC Working
> Group"
>
> On Tue, Jul 21, 2020 at 9:06 PM Yoav Weiss <yoav@yoav.ws> wrote:
>
>> +Ian Swett <ianswett@google.com> +Bence Béky <bnc@google.com>
>>
>> Google's Chrome and QUIC teams similarly intend to implement and support
>> reprioritization, for similar use-cases.
>> Upgrading images that are in or approaching the viewport as well as
>> downgrading the priority of large downloads both seem like important use
>> cases.
>> Video streaming use-cases where either quality-tier change or user
>> actions result in download changes also seem worthwhile (although I'm not
>> sure if cancellation can't handle some of those).
>>
>> Another use-case I heard from folks is that of JS reprioritization as a
>> result of user-actions: scripts that large apps want to download in
>> low-priority can become critical as a result of a user-action that needs
>> them. Being able to reprioritize can significantly impact such apps'
>> responsiveness.
>>
>> On Tue, Jul 21, 2020 at 7:59 PM Eric Kinnear <ekinnear@apple.com> wrote:
>>
>>> Hi all,
>>>
>>> We (both in Safari and in URLSession for general HTTP usage on Apple
>>> platforms) are quite excited about the new priorities document and the
>>> opportunities it brings to simplify and focus on information that’s
>>> strictly necessary to communicate between the client, server, and
>>> intermediaries.
>>>
>>> Specific to reprioritization, we have several cases where we use, and so
>>> far believe we need to continue to use, reprioritization (although ideas on
>>> workarounds are always welcome!).
>>>
>>> First, the Web download case that’s been discussed (and thanks Patrick
>>> for running some related experiments for web traffic!), where we use
>>> reprioritization to modify the incremental bit on resources.
>>>
>>> Second, when streaming HLS video, we prioritize the currently playing
>>> tier above the other tiers. We may have several requests outstanding for
>>> the next several segments of video, and when we switch up/down we need to
>>> be able to reprioritize those alternate tiers. Unfortunately, so far it’s
>>> looking like not being able to reprioritize these requests would prevent
>>> our implementation of the new priority scheme. For Low-Latency HLS, we
>>> certainly will need to use reprioritization if we’re to fit within the
>>> currently proposed priority tiers.
>>>
>>> Finally, a more generalized example. As we work to help customers and
>>> clients of the APIs we offer, we’ve found that many of our efforts to guide
>>> them towards appropriate prioritization of less important work at lower
>>> priorities is only enabled by the ability to raise that priority later when
>>> circumstances change.
>>>
>>> As a very contrived (but unfortunately close to real world) example,
>>> consider a case where we ask a client to de-prioritize loading of images in
>>> a list view that aren’t close to being scrolled into view by the user. If
>>> we can offer higher priority for those images once the user starts
>>> scrolling closer to having those items come into view, our clients are
>>> generally happy to initially load such images at lower priorities. However,
>>> if they’re stuck with that initial priority forever, they end up loading
>>> the entire set of images at a high priority *just in case* they might
>>> be eventually blocking render. A good bit of the time, that never happens,
>>> so we end up having everything at high priority when in reality we would
>>> rarely have needed to reprioritize the requests. And once everything’s at
>>> high priority, we no longer have the utility of the priority system at all.
>>>
>>> There are all sorts of ways to dissect that particular example, but the
>>> general response we’ve seen remains: folks are much more willing to fully
>>> utilize a prioritization system in the real world if they’re able to adjust
>>> the priorities that they assigned later on when they have more information
>>> or the circumstances change.
>>>
>>> Thanks,
>>> Eric
>>>
>>>
>>> Side note:
>>>
>>> For the document as a whole, we’ve gotten some feedback internally that
>>> it would be really nice if there were some (minimal, recommendation only)
>>> guidance as to how to respond to the priority signals when received. This
>>> wouldn’t be restrictive, as we’re really excited to experiment here and see
>>> what awesome results we can achieve, but having a baseline of “implement
>>> this as written and you’ll do *okay” *might be worth considering to
>>> increase the likelihood that we have a large group of generally-performant
>>> implementations.
>>>
>>> An example here would be if two requests of the same urgency arrive
>>> back-to-back, the first with the incremental bit set and the second
>>> without. What gets sent when? What do you do next if a third request
>>> arrives with the incremental bit also set before the first is complete?
>>> There are lots and lots of permutations, but a general approach of handling
>>> new items coming in is something that I think we’ve all been imagining
>>> during discussions, but we haven’t really written it down explicitly.
>>> Internally, as we discussed with some folks new to the topic, we discovered
>>> that our imaginations of what to do in cases like these didn’t actually
>>> align as well as we thought.
>>>
>>>
>>>
>>> On Jul 9, 2020, at 11:46 PM, Mark Nottingham <mnot@mnot.net> wrote:
>>>
>>> All,
>>>
>>> Thanks to everyone for their efforts so far.
>>>
>>> There's one other aspect that the we think it'd be helpful to get a
>>> sense of -- what the implementer intent is regarding reprioritisation.
>>>
>>> In particular, it'd be very helpful to have an indication from each
>>> implementation -- in user agents as well as servers (including
>>> intermediaries) -- as to how likely they are to produce/consume
>>> reprioritisations if specified.
>>>
>>> Note that's per-implementation, *not* per-person, so please coordinate
>>> if your implementation has multiple participants here.
>>>
>>> Responding to this e-mail is fine.
>>>
>>> Cheers,
>>>
>>>
>>> On 7 Jul 2020, at 7:50 am, Lucas Pardue <lucaspardue.24.7@gmail.com>
>>> wrote:
>>>
>>> Hi Patrick,
>>>
>>> Thanks for running this experiment and presenting the data back to the
>>> group.
>>>
>>> Also thanks to the Chrome folk for enabling the disabling flag.
>>>
>>> Cheers
>>> Lucas
>>>
>>>
>>> On Mon, 6 Jul 2020, 21:19 Patrick Meenan, <patmeenan@gmail.com> wrote:
>>> Sorry about the delay, just gathered the results.  The full raw results
>>> are here.  It looks like the impact dropped quite a bit across the full 25k
>>> URLs but looking at individual tests the impact is quite dramatic when it
>>> does impact (and it does exactly what we'd expect it to do for those
>>> outlier cases).
>>>
>>> The 95th percentile numbers tend to be the more interesting ones and in
>>> the data set, reprioritization enabled is the control and disabled is the
>>> experiment so positive changes means disabling reprioritization is that
>>> much slower.
>>>
>>> Largest Contentful Paint: 4% slower without reprioritization
>>> Speed Index: 2.75% slower without reprioritization
>>> Dom Content Loaded: 1.3% faster without reprioritization
>>>
>>> This is pretty much (directionally) what we'd expect since
>>> reprioritization boosts the priority of visible images (LPC/Speed Index)
>>> above late-body scripts (DCL). It's particularly dramatic for pages that
>>> use background images for any part of the page because they are discovered
>>> after all other resources and would normally be scheduled after all other
>>> scripts and inline images but if they are visible in the viewport the
>>> reprioritization helps them load much sooner.
>>>
>>> Looking at a few examples of the extreme cases:
>>>
>>> https://www.thehelm.co/ - (Filmstrip) - The main background image in
>>> the interstitial loads at < 10s vs 90s without reprioritization
>>> https://blog.nerdfactory.ai/ - (Filmstrip) - The background image for
>>> the main content loads at <5s vs 70s without reprioritization. No cost to
>>> DCL, just prioritized ahead of not-visible images.
>>> https://events.nuix.com/ - (Filmstrip) - Another hero background image
>>> (detecting a theme?) loads at 10s vs 60s
>>>
>>> Looking at a few of the bigger DCL regressions:
>>>
>>> https://oaklandcitychurch.org/ - (Filmstrip) - DCL got much slower (11s
>>> -> 33s) as a direct result of the background image moving from 30s to 10s
>>> (the pop-up interstitial was delayed along with the scripts that control
>>> it).
>>>
>>> For the specific case that most of these tests exposed (background image
>>> discovered late by CSS) it is theoretically possible for Chrome to detect
>>> the position before making the initial request (since it is only discovered
>>> at layout anyway) but that wouldn't help any of the more dynamic cases like
>>> when a user scrolls a page or a carousel rotates and what is on screen
>>> changes dynamically.
>>>
>>> I'm still of the pretty strong opinion that we need reprioritization but
>>> the web won't necessarily break without it and sites (and browsers) may be
>>> able to minimize the impact of not being able to reprioritize (though that
>>> might involve holding back requests and prioritizing locally like Chrome
>>> does for slow HTTP/2 connections).
>>>
>>>
>>> On Sat, Jun 20, 2020 at 10:17 AM Patrick Meenan <patmeenan@gmail.com>
>>> wrote:
>>> An early read on Yoav's Canary test is that most metrics are neutral but
>>> Largest Contentful Paint degrades ~6.8% on average and 12% at the 95th
>>> percentile without reprioritization and Speed Index degrades 2.6% on
>>> average and 5.4% at the 95th percentile. This is not entirely unexpected
>>> because the main use case for reprioritization in Chrome right now is
>>> boosting the priority of visible images after layout is done.
>>>
>>> We'll see if it holds after the full test is complete. The early read is
>>> from 3,000 of the 25,000 URLs that we are testing (all https hosted on
>>> Fastly for simplicity since we know it handles HTTP/2 reprioritization
>>> correctly).  The tests are all run at "3G Fast" speeds with desktop pages
>>> to maximize the liklihood that there will be time for reprioritization to
>>> happen.  I'll provide the full raw data as well as summary results when the
>>> test is complete (at least another week, maybe 2).
>>>
>>> On Wed, Jun 17, 2020 at 5:43 AM Yoav Weiss <yoav@yoav.ws> wrote:
>>>
>>>
>>> On Wed, Jun 17, 2020 at 9:55 AM Kazuho Oku <kazuhooku@gmail.com> wrote:
>>>
>>>
>>> 2020年6月11日(木) 6:46 Kinuko Yasuda <kinuko@chromium.org>:
>>> (Sorry, sent it too soon...)
>>>
>>> On Thu, Jun 11, 2020 at 6:12 AM Kinuko Yasuda <kinuko@chromium.org>
>>> wrote:
>>> Hi all,
>>>
>>> Reg: reprioritization benefit I can share some recent data for Chrome.
>>> For the two cases that are currently discussed I'm actually not fully sure
>>> about its benefit.
>>>
>>> For the renderer-triggered image reprioritization cases: this is a bit
>>> interesting one, we recently found two things:
>>> - Delaying to start low-prio requests could often work better (partly
>>> because of server-side handling) than re-prioritizing while inflight
>>> - In-lab measurements (tested with top 10k real sites, both on Mobile
>>> and Desktop) showed that removing in-flight re-prioritization doesn't
>>> impact page load performance a lot
>>>
>>> Let me stress though that testing this with servers that can properly
>>> handle reprioritization could change the landscape, and again this isn't
>>> really capturing how it affects long-lived request cases, or cases where
>>> tabs go foreground & background while loading, so for now I'm not very
>>> motivated to remove the reprioritization feature either.
>>>
>>> Hi Kinuko,
>>>
>>> Thank you for sharing your data. I feel a bit sad that reprioritization
>>> isn't showing much benefit at the moment. I tend to agree that we are
>>> likely to see different results between server implementations and HTTP
>>> versions being used. The effectiveness of reprioritization depends on the
>>> depth of the send buffer (after prioritization decision is made), at least
>>> to certain extent.
>>>
>>> FWIW, I added a flag to turn off Chromium's H2 request prioritization. I
>>> believe +Pat Meenan is currently running tests with and without this flag a
>>> list of servers we estimate is likely to handle them well.
>>>
>>>
>>>
>>> I suspect this is maybe because server-side handling is not always
>>> perfect and most of requests on the web are short-lived, and this may not
>>> be true for the cases where long-running requests matter.  I don't have
>>> data for whether may impact background / foreground cases (e.g. Chrome
>>> tries to lower priorities when tabs become background)
>>>
>>> For download cases, Chrome always starts a new download with a low
>>> priority (even if it has started as a navigation), so reprioritization
>>> doesn't happen.
>>>
>>> Kinuko
>>>
>>>
>>>
>>>
>>>
>>>
>>> On Wed, Jun 10, 2020 at 1:21 AM Lucas Pardue <lucaspardue.24.7@gmail.com>
>>> wrote:
>>> On Tue, Jun 9, 2020 at 4:27 PM Patrick Meenan <patmeenan@gmail.com>
>>> wrote:
>>> Eric's download example is a great one for exposing the risks that would
>>> come for an implementation that supported prioritization but not
>>> reprioritization.
>>>
>>> Take the trivial example of an anchor link that links to a download
>>> (say, a 200MB installer of some kind):
>>> - When the user clicks on the link, the browser assumes it is doing a
>>> navigation and issues the request with the "HTML" priority (relatively
>>> high, possibly non-incremental
>>> - When the response starts coming back, it has the content-disposition
>>> to download to a file.
>>> - At this point, the 200MB download will block every other
>>> lower-priority request on the same connection (or possibly navigation if it
>>> is non-incremental)
>>> - The user clicks on another page on the same site and gets nothing or a
>>> broken experience until the 200MB download completes
>>>
>>> Without reprioritization the browser will effectively have to burn the
>>> existing QUIC connection and issue any requests on a new connection (and
>>> repeat for each new download).
>>>
>>> Implementing prioritization without reprioritization in this case is
>>> worse than having no prioritization support at all.
>>>
>>> Thanks Eric for presenting this case, and Patrick for breaking it down.
>>> That does seem like a pretty bad outcome.
>>>
>>> Is this a good candidate for a test case? IIUC correctly the problem
>>> might occur today with HTTP/2 depending on how exclusive priorities are
>>> used. I'm curious if browsers can share any more information about what
>>> they do already. How does Firefox manage such a resource with it's priority
>>> groups?
>>>
>>> Cheers
>>> Lucas
>>>
>>>
>>>
>>> --
>>> Kazuho Oku
>>>
>>>
>>> --
>>> Mark Nottingham   https://www.mnot.net/
>>>
>>>
>>>
>>>