Re: [Tsv-art] Tsvart last call review of draft-ietf-httpbis-priority-10

[Lucas Pardue <[e-mail]>]

Hey Bob,

Responses inline (tagged [LP] because my mail client decided to be weird
today)

On Tue, Dec 21, 2021 at 12:16 PM Bob Briscoe <[e-mail]> wrote:

> Lucas,
>
> Thanks. Responses tagged [BB] inline where necessary. No comment =
> agreed/understood.
>
> Although I wrote this on 10 Dec, I held back from sending in order to
> think about fairness, but then forgot to finish it. I hope the extra time
> stewing has been worth it. I haven't gone back to all the issue trackers
> since then, except the one on fairness.
>
> Altho' you ack'd receipt of my updated review, I notice you've responded
> to the original incorrect one.
> So I've added back in my corrections and clarifications, tagged
> [BB-follow-up], because they toned down some of my snottiness, which I was
> trying to retract.
>
>
> On 09/12/2021 15:43, Lucas Pardue wrote:
>
> Hi Bob,
>
> Thanks (again) for the review.
>
> We've been working through the editorial issues and have cut release -11
> to address the ones we agreed with. Responses to the technical issues you
> presented are in-line below:
>
> On Tue, Nov 30, 2021 at 12:22 AM Bob Briscoe via Datatracker <
> [e-mail]> wrote:
>
>> Reviewer: Bob Briscoe
>> Review result: Almost Ready
>>
>> This document has been reviewed as part of the transport area review
>> team's
>> ongoing effort to review key IETF documents. These comments were written
>> primarily for the transport area directors, but are copied to the
>> document's
>> authors and WG to allow them to address any issues raised and also to the
>> IETF
>> discussion list for information.
>>
>> When done at the time of IETF Last Call, the authors should consider this
>> review as part of the last-call comments they receive. Please always CC
>> [e-mail] if you reply to or forward this review.
>>
>> Version reviewed:
>> Originally draft-ietf-httpbis-priority-09, but also checked -10 diff.
>>
>> ==Summary==
>> The move to an e2e request and hop-by-hop response is a good one.
>>
>> I'm not sure how ready this is, until I see how the authors respond to my
>> questions about the interaction model and whether the client can say "I
>> dunno,
>> you tell me" to the server (T#4c, T#5a and T#9a).
>>
>> I think all of my other points are 'just' holes in the coverage of each
>> aspect
>> of the protocol, but some will be quite involved to fill. There's a lot of
>> vagueness still to be tied down, I'm afraid.
>>
>> Three sets of comments below:
>> * Gaps (G#): 2
>> * Technical points or suggested modifications to normative text (T#) 13
>> * Editorial points (E#) 29
>>
>> ______________________________________________________________
>> ==Gaps==
>>
>> G#1 Implementation status section?
>>
>> This review would have really benefited from an implementation status
>> section.
>> You will see I got suspicious that some of the sections had been written
>> without the benefit of any implementation or operational experience. While
>> others seemed stronger. If the implementation status had been written up,
>> I
>> wouldn't have had to guess.
>>
>> Here's my guesses at what has been implemented, given the waffle factor
>> of the
>> relevant sections ;)
>>
>> [BB-follow-up]: Sorry, 'waffle factor' was a bit derogatory. I meant that
>> the sections that I have identified below as 'less mature' tended not to
>> say anything significant, while apparently looking as if they were saying
>> something significant.
>>
>>
>> * e2e priority protocol protocol handlers and schedulers: most mature
>>
>> * intermediary priority protocol handlers and schedulers: not so mature
>>
>> * automated priority header generation, APIs: not so mature
>>
>> * priority of server push, retransmissions, probes: just ideas in theory?
>>
>> * investigation of deadlocks, loops, etc: early days.
>>
>
> The short answer is that HTTP prioritization is much of an optional thing.
> Signals are hints to the server in charge of the multiplexed connection.
> Servers have a self-interest in serving requests in a timely manner
> balanced against all the other needs like resource usage, DoS avoidance and
> so on. The sections that describe scheduling are some basic recommendations
> that will let clients have some expectations how responses would get
> prioritized should the stars align. But this is not a very specific
> algorithm that all implementations will follow exactly because rarely do
> stars align. The sections attempt to spell out the considerations arising
> from the protocols related to this draft. We believe the editorial changes
> made in -11 make it clearer where text is considerations more than
> authoritative direction on what to do.
>
>
> [BB] I think you might have missed my point. I meant a section telling us
> what degree of implementation there has been of each aspect (to be removed
> before publication as an RFC). I should have referenced BCP 205 [RFC7942].
>
> Reason: Unless readers put in effort to track what everyone else is doing,
> they have no idea whether the various aspects of this draft are based on
> arm-waving or real operational experience. For anyone who is implementing
> themselves, that colours how much weight they should give to the advice
> relative to their own experience.
>

[LP] The nature of this spec means that the majority of the work - server
scheduling - is delegated solely to server implementation detail. The
considerations presented hold whether people choose to take them into
account or not. How large an effect the considerations have on any single
implementation depends on the nature of the implementation, the deployment
environment, the runtime conditions under which any single connection is
operating, etc.

Given where we are at in the document cycle, adding such a section would
just delay progress and I'm not convinced it will improve this document for
the target audience.

>
>
>
>
>> G#2 Performance evaluation?
>>
>> Priorities are about improving performance. This is a stds track draft
>> about a
>> core IETF protocol. But there is nothing in this draft pointing to any
>> studies
>> that quantify how much performance is improved (or not) by the different
>> aspects of the protocol. Ideally there would be a study comparing the
>> HTTP/2
>> priority approach with this one. Is that because the studies don't exist,
>> or
>> just an omission?
>>
>> [BB-follow-up]: I was wrong to say there is nothing pointing to any
>> studies, just that the studies pointed to haven't been cited. I should have
>> just asked for references to be provided at the end of §2, after:
>>
>>     "Multiple experiments from independent research..."
>>
>>
> We simply overlooked citing them, which was noted in other reviews. We
> added links to work from Robin Marx and Pat Meenan that motivated this
> document and it's design features.
>
>
>>
>> ______________________________________________________________
>> ==Technical Points and Modifications to Normative Statements==
>>
>> §2.  Motivation for Replacing RFC 7540 Priorities
>>
>> T#2a) Grass is always greener on the other side of the fence?
>>
>>    RFC 7540 priority is expressed relative to other requests on the same
>>    connection.  Many requests are generated without knowledge of how
>>    other requests might share a connection, which makes this difficult
>>    to use reliably
>>
>> This says why relative values were difficult, but it doesn't say why or
>> whether
>> absolute values will be better. Is there lots of experience of absolute
>> values
>> being sufficient and easier to use? or any cases where absolute values
>> might be
>> insufficient? It seems obvious that at run-time you might hit the end of
>> the
>> number space, i.e. have to pile up objects on the same urgency value at
>> the
>> edge of the space when you really wanted some objects to have higher (or
>> lower)
>> urgency. There is a mention of the possibility of creating an extension
>> with
>> intermediate urgency values, but what does an implementation do when it
>> hits
>> this problem in the middle of a session? It can't quickly jump out of
>> run-time,
>> convene a design team to get a new scheme extension agreed then go back to
>> run-time and complete the session.
>>
>
> Dependencies between the requests in a connection are a property of the
> connection. This poses challenges to acting on priority signals when
> passing them to components/nodes that are not part of the connection.
> That's friction against some ways that HTTP commonly operates in practice.
> For example, intermediaries or servers that split or coalesce requests from
> different connections. The urgency and incremental parameters have been
> deemed sufficient for a web browsing use case, which was our focus, without
> being too fine grained and hard for developers to reason about. If other
> use cases encounter limitations or problems with this scheme, I do
> encourage them to bring that back to the HTTP WG so we can consider work on
> extensions that address them.
>
>
>>
>> T#2b) Motivation for replacing 7540 included removal of dependencies?
>>
>> The Security Considerations says that one of the motivations for
>> replacing 7540
>> was that "Extensible priorities does not use dependencies, which avoids
>> these
>> [resource loop DoS attack] issues." Draft-09 listed this as one of the
>> motivations in §2, but in draft-10 it has been removed from §2. If it is
>> still
>> a motivation, it ought to be listed in §2, not just in Security
>> Considerations.
>>
>> Security Considerations seems to be in the commonly used style of just a
>> list
>> of pointers to other parts of the draft. So it would be consistent to say
>> this
>> in §2 not just in Security Considerations, which even says
>> "Considerations are
>> presented to implementations, describing how..." as though the details are
>> elsewhere in the document.
>>
>> Whatever, given this seems to have been an important motivation, please
>> try to
>> describe this issue in a self-contained way, rather than talking
>> obliquely in a
>> way that requires the reader to refer to the CERT advisory (e.g. "...some
>> types
>> of problems...").
>>
>
> The text was written before the 7540bis activity started. We’ve been
> shifting bits and pieces of RFC 7540 problems to that venue. This one has
> been removed since it seems like a distraction for the topic of this
> document.
>
>
> [BB] I notice it's now removed from Sec Consids as well.
>
> That's fine, but BTW, it wasn't a distraction for me. It seemed to be one
> of the stronger reasons for moving away from dependencies, because (in many
> cases) I can't understand how priorities can be implemented without knowing
> the dependencies (see T#9a).
>

[LP] Ack. We think the text in the latest version now pays enough detail to
RFC 7540 without treading on 7540bis toes.


>
>
>
>>
>> §2.1.  Disabling RFC 7540 Priorities
>>
>> T#2c) Incremental deployment
>>
>> Two perhaps obvious but unstated things ought to be stated:
>> i) An HTTP session will always _function_ even if all priority
>> information is
>> ignored; it just might perform badly.
>>
>
> [BB] I think you've overlooked this point.
>
> In your very first response above - about implementation status - you say
> "The short answer is that HTTP prioritization is much of an optional
> thing." The point I was making here is that I don't believe the draft ever
> actually explicitly says the obvious - that HTTP will function without any
> priorities at all. And even if the lack of MUSTs ultimately implies
> everything is optional, it would be worth making that clear from the outset
> in the Introduction.
>
> BTW, when I asked for the refs to evaluations, I was trying to establish
> whether the gain from prioritization is turning out to be worth the pain of
> implementing it and driving it at run-time.
>

[LP] HTTP/2 and HTTP/3 (and QUIC) are the protocols that define
mulitplexing and concurrency and they contain comments roughly equating to
prioritization is good, no prioritization is bad. Neither 7540bis, RFC 9000
or draft-ietf-quic-http-34 define prioritiy signals.

Can you elaborate on what problem is solved by adding the statement you
suggest to this document?


>
>> ii) The semantics of the SETTINGS_NO_RFC7540_PRIORITIES setting is
>> intended to
>> apply to both directions (if it is?). When it says
>>   "A server that receives SETTINGS_NO_RFC7540_PRIORITIES with a value of
>> 1 MUST
>>   ignore HTTP/2 priority signals."
>> I assume after "MUST ignore" it intends to add "...and MUST NOT send...".
>>
>
> [BB] I don't think you've addressed this point below or in the linked
> issue you've opened. Altho there is no appetite in the WG for defining what
> 7540 should have said, surely it would be worth closing off any uncertainty
> for the future, given all the opinions in response to this point were
> purely based on what they /think/ has not been implemented. It would do no
> harm to just add that a 7540bis server in receipt of
> SETTINGS_NO_RFC7540_PRIORITIES =1 MUST NOT send HTTP/2 priority messages
> (as well as ignoring incoming).
>

[LP] Other documents talk in terms of servers operating on client signals.
They don't mention server->client PRIORITY frames. So in effect, the signal
is always ignored anyway. Trying to talk to that direction of flow of the
signal here seems like it would add complication to the document and I
think a distraction.

If there is some use of server->client signals, then the HTTP WG is not
aware of it. If the people that use server->client signals feel like they
need to disable them, then then I think they are best placed to document
the concerns for that.


> I assume this is stated as "server MUST ignore" rather than a protocol
>> error,
>> because the HTTP/2 priority signals might have come from an intermediary
>> that
>> doesn't understand the SETTINGS_NO_RFC7540_PRIORITIES setting.
>>
>
> [BB] This was a question really. The subsequent points depend on the
> answer.
>
> If the answer is that HTTP/2 priority signals could never arrive at the
> server from an intermediary, and if a client says it won't send a certain
> header, but subsequently the server receives one anyway, surely that is a
> strong case for protocol error.
>

>
>> Also, it is surely a protocol error if one endpoint sets
>> SETTINGS_NO_RFC7540_PRIORITIES to the opposite of the other.
>
>
> [BB] I don't think you've addressed this point either.
> What is meant to happen if
> * an HTTPbis client sends SETTINGS_NO_RFC7540_PRIORITIES = 1
> * the HTTPbis server sends SETTINGS_NO_RFC7540_PRIORITIES = 0
> ?
>
> This means the client is "not using" HTTP/2 priority signals and the
> server says I "won't ignore" HTTP/2 priority signals I receive. Altho' that
> would lead to no protocol problems, doesn't it imply that the server is
> somehow broken? I agree that the old HTTP/2 priority signals are hints, but
> this SETTINGS message is not a hint. It's a configuration setting
> introduced in HTTPbis. So isn't it worth defining it fully?
>
>
> Or if a node sends
>> a header after it has said it won't.
>>
>
> [BB] Again, this relates to the new SETTINGS message, not the fact that
> 7540 priority messages are hints. Surely errors like this need to be caught
> (they could e.g. be due to a load balancer switching to an endpoint with a
> different config).
>

[LP] Responding to the three points as one:

Given the current constrains and how loosely coupled things, I'm not sure
what problem is solved by a protocol error.

H2 says - "Settings are not negotiated; they describe characteristics of
the sending peer, which are used by the receiving peer.". And that governs
a lot of what this section can do.

The initial value of SETTINGS_NO_RFC7540_PRIORITIES is 0. A client that
understands the setting will assume 0 when SETTINGS arrive and
SETTINGS_NO_RFC7540_PRIORITIES is not present. So the client=1, server=0
case exists already implicitly.

One likely deployment model for the Extensible priorities scheme is for
HTTP applications to add the priority header on request messages, in
ignorance of whatever the HTTP stack underneath supports or is doing. That
lets them write version-independent logic. Being too strict with settings
rules risks adding hard-to-diagnose errors for such applications.


> If a client sets SETTINGS_NO_RFC7540_PRIORITIES to 1, but the server
> doesn't
> understand this setting, and later sends HTTP/2 priority signals (perhaps
> in
> response to an intermediary), what happens?
>
[As I pointed out in my review of
> RFC7540 Priorities (when it was a draft but after IESG approval), it wasn't
> clear whether priority messages were only sent in the C-S direction, or
> also
> the reverse. I didn't receive a reply on that point and the RFC is still
> not
> clear.
> https://lists.w3.org/Archives/Public/ietf-http-wg/2015JanMar/0529.html ]
>
> [BB-follow-up]: That wasn't intended to sound like sour grapes. In the
> opening sentence of my review of http/2 I had said I was aware that my
> (long) review had arrived after http/2 had been approved by the IESG. My
> review was unsolicited. And I guess the last thing anyone wants to do when
> their draft has just been approved is answer a load of points about
> something they are not going to change. But I guess it should have been
> handled eventually, if only for errata.
>
>
Francesca also mentioned this in the AD review. RFC750 seems to have
allowed server-to-client signals on the wire but didn’t specify what at all
anyone should do with them. I’m not aware of any cases of this signal being
used in the wild. I created an issue on
https://github.com/httpwg/http2-spec/issues/1000 there's further discussion
there. 7540bis deprecates stream prioritization - all that is left is the
remnant of bits on the wire that remain for wire compatibility.

SETTINGS_NO_RFC7540 is an optimization related to carriage, processing and
application of signals. Because these signals are only a hint, as long as
they have a valid wire format there is no need for protocol errors. Since
we are focused on C->S direction of signal, and the world never defined
what RFC7540 S->C signals really meant, our document doesn’t benefit from
trying to speak about them. So we define our terms of use in the Notational
Conventions and stick to them. The document consistently uses this term,


[BB] Above, I have highlighted again those of my points that were about how
> the SETTINGS message itself (not the 7540 hints) is not fully tied down.
>



>
> §4.  Priority Parameters
>
> T#4a) Vagueness permeates what intermediaries do in this draft
>
>    Intermediaries can consume and produce priority signals in a
>    ...PRIORITY_UPDATE frame or Priority header field.
>    ...Replacing or adding a Priority header field overrides
>    any signal from a client and can affect prioritization for all
>    subsequent recipients.
>
> * Do intermediaries really both consume and produce priority signals.
> Always?
> In both directions? What does 'consume' mean (absorb and not forward, or
> read
> and forward)?
>

They can according to HTTP Semantics.


> * Can they really use either type of frame? Always?
>

They can, it depends on the versions of HTTP being used on the upstream or
downstream.


> * How does adding a priority header override any signal from a client? Or
> is it
> only replacing that overrides?
>
> My later comment asking for a more precise statement of the protocol's
> interaction model ought to resolve these issues as well.
>

it all depends on the model about how HTTP intermediaries convert between
versions. Frames are connection-level and headers might be e2e or
hop-by-hop.  This document shouldn’t litigate any more than it does on the
matter.


> T#4b) Really only C-S direction?
>
>     PRIORITY_UPDATE frame preserves the signal from the client, but...
>
>     ...overrides any signal from a client...
>
> Also used for S-C direction?
>
> Given this part of the draft seems to have been written solely about the
> C-S
> direction, perhaps it would be better to admit that is a good way to
> structure
> the draft with C-S first. Then add another section about S-C, and perhaps
> another about S-Int. The alternative of adding to all the definitions to
> cover
> all directions and interactions, might become incomprehensible.
>

This is different from HTTP/2 PRIORITY frames (as mentioned above). By
definition PRIORITY_UPDATE is only allowed in the C->S direction, which
eliminates the need to document the reverse direction.


[BB] Then this becomes just an editorial point, that it would be useful to
say explicitly at the start of §4.

CURRENT:

   PRIORITY_UPDATE frames (Section 7.1 and Section 7.2) are used by
   clients...

PROPOSED

   PRIORITY_UPDATE frames (Section 7.1 and Section 7.2) are sent by
   clients (not by servers)...


There seems to be a pattern developing (RFC7540 and now here) where the
people who write the spec know what the interaction model is and they
assume everyone else does, which sometimes becomes revealed implicitly as
the reader gets into the details, rather like a game of Cluedo. It would be
really, really helpful if you just stated it early on.


The useful change is to clarify it's the client that sends the frames. I
made https://github.com/httpwg/http-extensions/pull/1866 to do that.
Stating client (not server), server (not client) is not a style that I
think would improve the readability of this draft.



>
> §4.2.  Incremental
>
> T#4c) Client doesn't always have prerequisite info to set incremental
> parameter
>
> There will surely be cases where the MIME type of the response (and
> therefore
> whether the client can render it incrementally) is not known, or cannot be
> guessed by the client when it requests a resource, or when it starts
> content
> negotiation? For instance, the client might have listed MIME types in its
> Accept list, some of which are incremental, and some not.
>
> [BB-follow-up]: I intended to add that I suspect in common cases the
> client's guess of what MIME type is likely to served will be sufficient and
> correct. This comment was meant to be about the remainder of cases.
>
>
>
> The server can't override a client 'not incremental' message by stating
> that
> the MIME type it has served is incremental. Because, when the client says
> 'not
> incremental', that is intended to state the capability of the client, not
> the
> format of the resource.
>
> Perhaps the HTML that gave the client the hyperlink that was selected to
> get
> the resource could also include a tag giving the MIME type of the
> hyperlinked
> resource? Or perhaps the idea is that the client has to send a
> PRIORITY_UPDATE
> once it knows the MIME type (by which time it might be too late)?
>

That's a fair point. Unfortunately, there will be cases where parties lack
all of the information that could lead to perfect prioritization. Client
priority signals are only a hint. Servers can and will do whatever they
like, including serving the response in a way that does not follow the
recommendations we provide for handling the incremental parameter in
Section 10. There’s lot of additional means, outside of this specification,
that clients and servers can use to augment their understanding of
priority. There is no need to enumerate them in this document.


[BB] Strictly then, the incremental parameter is a property of each MIME
type in the Accept list, rather than a property of the request as a whole.
But I guess it is a reasonable assumption that most of the commonly used
selections of alternative MIME types will all share the same incremental
property.

[LP] Ack. I think what we have is ample enough.



> §5.  The Priority HTTP Header Field
>
> T#5a) Interaction model: an example or mandatory?
>
> It would help to start by explaining (perhaps in the Intro, rather than §5)
> whether a priority message about a response can be initiated by a server or
> intermediary if there was not a priority field attached to the request
> from the
> client. I believe the draft intends this not to be possible, although this
> is
> not stated normatively anywhere, and I don't know why such a restriction
> would
> be imposed.
>
> Actually, I believe it is essential that the protocol allows the server to
> initiate priority messages, as absence of a message is currently the only
> way
> for the client to say "I have no idea, you decide". Otherwise, if the
> server is
> only allowed to follow the client, when the server knows the best order to
> serve the objects (which I believe is often the case), the client still
> has to
> request non-incremental objects in some order or other, and give them some
> priority or other. So the server doesn't know whether the client actually
> knows
> what it is doing, or whether it is just making up an ordering because it
> has
> to, even tho' it has no clue.
>
> [BB-follow-up]: When I said "which I believe is often the case" I didn't
> intend to mean the majority of cases. I meant "not an insignificant number".
> I'm sure the client can have a crack at how to set the priorities in
> common cases. Which can then be modified by the server if necessary.
>
> The point I was trying to highlight was that the client has no way to tell
> the server how to distinguish between "I'm certain about these priorities"
> and "I'm not really sure, but the protocol requires me to look as if I'm
> sure, even if I just leave everything as default priority."
> I think I explained it better under T#9a.
>
>
> Alternatively, could the client send a Priority header with no parameters?
> This
> would indicate that the client wants the server to prioritize, and to
> allow the
> server to tell intermediaries what to prioritize. (For more about clueless
> clients, see T#9a) "Client scheduling".)
>
> The abstract gives the only outline of the interaction model, but it's not
> clear whether this is just an example of common usage, or the only possible
> model.
>
> §5 just says the priority field can be used "when a request or response is
> issued". It goes on to state that the priority field is an e2e signal, but
> then
> in the next sentence talks about how intermediaries can combine priority
> info
> from client requests and server responses (which reflects what §8 says as
> well). So "e2e" is clearly an over-simplification. I think it's e2e in one
> direction but hop-by-hop in the other (supported by the description in the
> abstract), ie. client -> server -> intermediary/ies -> client. It's also
> possible that intermediaries are intended to (or at least allowed to) read
> but
> do not alter the messages in the C-S direction, otherwise, what would they
> 'combine' with the priority field coming from the other direction?
>
> Whatever, the interaction model(s) is never stated precisely. I've
> classified
> this as a technical point, not just editorial, because I couldn't really
> assess
> the completeness of many other technical details of the draft without
> knowing
> the answer to this fundamental question.
>

If this scheme is implemented at the server, then all requests are treated
as if they have an associated priority signal. This can be a Priority
header field (Note the serialization rules for Structured Fields Dictionary
in Section 3.2 of RFC 8941 - sending an empty header is not permitted) or a
PRIORITY_UPDATE frame. Omission of signals, or omission values in signals,
invokes default value priority parameters urgency=3, incremental=false. A
server is therefore able to determine the client’s view of the priority.

Editorial changes now in Section 10 should make it clearer that an HTTP
server’s job is to respond in a timely manner. And it always has to decide
how to use finite resources to do so. Clients can hint as some preference
but if they don’t know or don’t care, it's basically delegating the
responsibility to the server.

The purpose of the Priority header in responses is to allow origin servers
(detached from the intermediaries’ connection to the client) the ability to
also provide hints about prioritization.

The interaction model is described throughout the document, with a gist in
the intro. Duplicating details into the intro does not seem beneficial.


[BB] This is a major problem for the reader. All the way through, I'm
> having to build a set of possibilities of what might be meant, then
> gradually remove them one at a time, which is why I likened it to a game of
> Cluedo.
>
> An interaction model is never about 'details'. I'm not asking for details.
> I'm asking that, if the model described in the abstract is the only model,
> it says so. And if it's only an example, it says so. It is a common mistake
> for the authors of a protocol to omit description of the big picture,
> probably because there has never been any other possibility of a different
> big picture in /their/ minds.
>

[LP] In that case, I don't really follow what you're asking for. We note
the actors and the messages between them. Just adding a sentence to Section
1 of paragraph 6 that says "the only thing we've defined applies to the
only this we've defined" seems like waffle to me. Maybe you have a better
concrete suggestion?



>
> T#5b) Normative 'cannot'?
>
>    Clients cannot interpret the
>    appearance or omission of a Priority response header as
>    acknowledgement that any prioritization has occurred.
>
> Was this intended to say 'Clents MUST NOT interpret...'?
>

Signals by design are just a hint. They can never be trusted and this
sentence highlights that fact.


> T#5c) Nothing said about caching and priority
>
> The paragraph about caching and priority just ends having talked a bit
> about
> caching but not about priority. It left me none the wiser about what a
> cache
> ought to store about priority with the response. §13.8 talks about fairness
> between multiple live connections in the presence of coalescing. But
> doesn't
> the discussion of caching and priority here need to talk about what
> must/should/may be stored about priority in a cache for later connections.
> Even
> if it's implementation dependent, wouldn't it be worth a brief discussion
> (as
> in the 2 paras below).
>
> The priority of a response is the outcome of an interaction between the
> client's original (e2e) priority combined with the server's logic about the
> resource. If only the priority outcome is stored, then when another request
> arrives at the cache from a different client, there will be no record of
> the
> original client's priority. So the  cache will not know what client
> priority
> led to the priority stored with the response. And it will not know whether
> the
> current client priority is the same or different.
>
> On the other hand, if the cache stores the original client priority with
> the
> response priority, then should it refer a request with a different (e2e)
> client
> priority to the server, then store the new pair of priorities with the
> original
> cached response? And I guess it could serve the request in parallel, rather
> than waiting for the server to tell it whether to serve the request
> urgently
> (!). This would probably scale reasonably well, given the likely small
> number
> of different client priorities. But who knows how it would scale if the
> parameter space is extended in future.
>

Answer supplied by Kazuho - As discussed in the last paragraph of section
5, CACHING defines if and how requests with different header field values
can be mapped to one response. If the capabilities provided by CACHING
(i.e. Vary) is too limited, then we should fix that as an extension to
CACHING (as have been previously proposed as draft-ietf-httpbis-key). In
practice, re Extensible Priorities, IMO, there aren't many sensible
combinations of urgency and incremental. Therefore, backend servers that
want to tune priority based on the value that the client sends can simply
send Vary: priority and call it a day.


[BB] I think my point has been missed. I'll try an example:
> Client A requests
>     priority u=4
> Server responds
>     priority u=2,
> which gets cached.
> Client B requests same object
>     priority u=4.
> Client C requests same object
>     priority u=0
>
> If requests B & C were forwarded to the origin, it would respond with
>     priority u=2    # for B
>     priority u=0    # for C
>
> However, even though the cached object has the same priority header that
> the origin server would give to Client B's request, it's different to that
> cached. And the cache cannot tell that B's request would match, but C's
> wouldn't.
>
> Vary doesn't help here, does it? At least not without storing the client
> request priority as well as the server response.
>

[LP] I'll let Kazuho respond to this one.



> §9.  Client Scheduling
>
> T#9a) Client doesn't have prerequisite info about content sizes and
> dependencies
>
> Consider a web page example with a load of non-incremental objects for the
> client to request. It doesn't know their sizes, and it doesn't know which
> ones
> might contain references to further objects to request. So it requests
> A,B,C,D.
> In retrospect, it turns out that C was huge, and D referred to further
> objects
> to download. How was the client to know it should have downloaded D before
> C?
>
> To be effective, a scheduler needs to know object sizes and which objects
> will
> identify further objects to be requested (dependencies).
> * Size is known by the server but not by the client, at least not until
> the headers
> at the start of the object arrive.
> * Dependencies are known by the server, but not by the client until an
> object
> starts to unfold.
>
> Why is the client made to choose the priorities of the responses? It
> doesn't
> know any of this vital metadata about these objects. It can guess from file
> types that JS and HTML probably ought to come first. But it knows little
> else.
>
> So, as I already said under my question T#5a) about the interaction model,
> the
> most important capability the client must have is the ability to say "I
> dunno,
> you decide". But that's the one thing this draft doesn't allow the client
> to do
> (at least I think it doesn't? see T#5a). For a list of non-incremental
> objects,
> even if the client gives all their requests the same urgency, it can't
> send all
> the requests at the same time - it has to choose which order to send them
> in,
> even if it has no clue. This tells the server to respond in that order OR
> to
> choose a different order. But the server doesn't know whether the client
> chose
> this order deliberately or just because it didn't know any better.
>
> Alternatively, there will need to be some way for the server to tell the
> client
> what to prioritize _before_ it sends its requests (e.g. using extensions
> to the
> HTML in a base HTML document).
>

As noted in response to T#4c, we are constrained by the capabilities of
information exchange that HTTP and its uses (such as the web) allows us.
This is no different a problem than existed for RFC 7540. Only a client
knows how it wants to use resources for which it has limited knowledge.


[BB] it's not the case that only a client knows how to use resources...
> That's the whole problem.
> The position is that B (the server) has to decide how to prioritize, but B
> knows info that A needs in order to tell B what to prioritize. And there's
> limited time for them to chat about it before prioritization becomes moot.
>
> So either B tells A the missing info, so A can tell B what to prioritize.
> Or A tells B its various strategies that depend on the info B holds. The
> latter is complex, the former would make more sense, but only if the
> protocol exchange had started earlier (!)
>
> I think the new priority approach is deliberately not fully solving this
> problem (unlike http/2 priorities), on the basis that httpbis can still add
> benefit for the current typical way web objects are structured. If you
> agree, I this is true, the spec ought to say it (nothing to be ashamed
> about).
>

[LP] RFC 7540 didn't have a solution to this. The information available to
clients is independent of the scheme used to signal a priority hint from
client to server.

In the "I dunno" case for RFC 7540, the defaults are invoked (non-exclusive
dependent on stream 0x0, weight 16) [1]. So all requests had an implicit or
explicit priority; an explicit prioritization to stream 0x0 with weight 16
is indistinguishable from an implicit priority.

I'm still unclear what you're asking for. Is it some short editorial text
to discuss client considerations for choosing the value of u or i
parameters?

If we use an HTML document as an example, the subresources have a
dependency chain that may or may not change while they get loaded. It’s
more likely that a client will request something with a particular priority
because of its type and usage in a given HTML document, rather than its
size (even if it knew that size). It’s going to be rare that a client
doesn't have an opinion - if the client doesn’t know, the defaults are
sufficient to let it reprioritize the request higher or lower if it finds
out that need once the response is back.

But really this all comes down to making all the actors aware of the
challenges and stating that priority signals are just hints in the decision
making. If a client finds that the server is making scheduling choices when
using defaults, then it is empowered to provide better signals.

The general problem is not solvable so I do not believe there is anything
more we can add to the document.


[BB] Apart from the fundamental problem of A&B having info in the wrong
> place (see above)... I believe the missing piece is an ability for the
> client to be able to distinguish between "I'm certain" and "I really don't
> know". As you explained earlier, "all requests are treated as if they have
> an associated priority signal", so the client has no way to express "I
> really don't know".
>
> I've said that multiple times now, and it seems to be falling on stony
> ground,  so I'll stop. Maybe I'm wrong, but I've said as strongly as I can
> that I think this will be a problem.
>

I think we can agree to disagree on this one. If it does really turn into a
problem, we can convert that into an opportunity to define a new parameter
for this scheme to address it. For example, imagine a new "confidence"
parameter between 0 and 100 that expresses the confidence of the priority
signal it was sent with.,



>
> §10.  Server Scheduling
>
> T#10a) Server push priority sounds like waffle
>
> [BB-follow-up]: When I said "which I believe is often the case" I didn't
> intend to mean the majority of cases. I meant "not an insignificant number".
> I'm sure the client can have a crack at how to set the priorities in
> common cases. Which can then be modified by the server if necessary.
>
> The point I was trying to highlight was that the client has no way to tell
> the server how to distinguish between "I'm certain about these priorities"
> and "I'm not really sure, but the protocol requires me to look as if I'm
> sure, even if I just leave everything as default priority."
> I think I explained it better under T#9a.
>
>
> The discussion of priority for server push seems to say "This might not
> work".
> If this conclusion is based on operational experience it ought to say so.
> And
> if it's not, it ought to say that it's just conjecture.
>

The topic of server push is not helped by the fact that it’s deployment
story, outside this I-D, is one of disappointment. But we’re stuck with
that feature in HTTP/2 and HTTP/3 and a server that chooses to use it while
implementing this scheme has to make some choices. There's no case where
push will fail but there are cases that could cause it to perform badly.
The text in paragraphs 10 and 11 provide considerations that a server that
does implement server push will have to make (because things are _always_
contending for resource). That said, I don’t think we need operational
experience to conclude that if you push things incorrectly, you could
really hurt what the client is trying to achieve.


[BB] This doesn't answer my question. The text is essentially blustering.
> Please just say that there is little operational experience to go on, if
> there isn't. Rather than trying to say something that sounds wise,... and
> failing.


[LP] I disagree. Server push is the initiation of a response based on a
request that a client never made. The server has to pick something. And we
know that what it picks will interplay with non-pushed resources. We spell
out the problem that if you don't consider these things, problems could
happen. We don't need operational experience to reason that guessing wrong
is suboptimal when picking scheduling using this scheme.



> §12.  Retransmission Scheduling
>
> T#12a) Sounds like more waffle
> Similarly, if retransmission scheduling and/or probe scheduling has limited
> operational experience or limited usefulness, it would be better to say so,
> rather than trying to sound authoritative without really saying anything.
>

Again, this is something a server has to do anyway and we want to present
the considerations at play. Our editorial change makes it sound less
authoritative by nixing the sentence containing “...its effectiveness can
be further enhanced…”.


>
> §13. Fairness
>
> T#13a) Please define fair. Seriously.
>
> A prerequisite question is, "What parameters does a server scheduler
> manipulate?" The text implies the server can only control the order in
> which it
> starts a response to each request, and whether responses start while other
> responses are in progress or wait for their completion. If so, I'm not sure
> what fairness could mean.
>
> Presumably a server can also determine the relative rate at which it sends
> different streams. And it could stall a stream to allow another absolute
> priority. In this context, fairness might mean instantaneously equal
> rates. But
> that's not fair if the objects are of different sizes.
>
> So we genuinely do need to see a definition of what fairness means here.
>

Good point, we addressed this in issue 1819 -
https://github.com/httpwg/http-extensions/issues/1819


[BB] I've read the issue, but I think it reveals unstated assumptions about
> what fair means. To understand what I mean by "equal rates are not fair if
> the objects are of different sizes", please see the description around Fig
> 1 in this discussion paper:
> https://bobbriscoe.net/projects/latency/per-flow_tr.pdf#page.3
>
> (The context is transport layer, not app-layer, but the arguments are the
> same. In the next response, I try to pull together app-layer and transport
> layer fairness).
>


> T#13b) Why not make scheduling decisions across different clients?
>
>    As a general guideline, a server SHOULD NOT use priority information
>    for making scheduling decisions across multiple connections, unless
>    it knows that those connections originate from the same client.
>
> Why does the IETF have anything to say about this? It's surely an operator
> policy decision.
>

We disagree a bit with this but we discused some more on issue 1820 -
https://github.com/httpwg/http-extensions/issues/1820


[BB] I've read 1820, and the new text in draft-11.

   Typically, HTTP implementations depend on the underlying transport to
   maintain fairness between connections competing for bandwidth.

Interesting. But you (the apps area) don't appear to know your own
> strength...
>
> When the endpoints set the non-incremental parameter on streams A,B,C,D,E
> , it is equivalent to giving B-E zero rate until A completes, then C-E zero
> rate until B completes, etc.
>
> So, I think most transport area people would say the opposite of the
> above; the transport layer handles relative rates of connections, but only
> in the absence of more conscious bandwidth scheduling driven from the app
> layer (or intervention by the network operator, which is the other
> stakeholders that might apply bandwidth policy, but it has very weak
> visibility and knowledge of the information it is handling).
>
> The transport area offers no meaningful guidance on app-layer fairness.
> For instance, the app-layer decides:
> a) which streams to hold back from being allowed to send (as just
> described above);
> b) which streams are collected together within one 5-tuple flow;
> c) how much data to send into each socket.
> So the app-layer determines the units of fairness that the transport layer
> has to work with. The transport area's only contribution has been
> rate-equality between flow IDs competing in a common bottleneck. So I
> suggest you consider transport layer 'fairness' as just an arbitrary
> default policy that is merely the easiest way to prevent starvation. But
> always remember that the app-layer can override the transport layer and
> impose starvation (e.g. with the non-incremental parameter, as described
> above).
>
> In the transport area, there have been some recent advance in
> understanding, allowing that rate equality between flow IDs only becomes
> more necessary as congestion rises. The corollary being that, when
> congestion is low, more inequality is OK. IOW rate-equality is most useful
> as a famine response.
>
>    When
>    HTTP requests are forwarded through intermediaries, progress made by
>    each connection originating from end clients can become different
>    over time, depending on how intermediaries coalesce or split requests
>    into backend connections.  This unfairness can expand if priority
>    signals are used.  Section 13.1 <https://datatracker.ietf.org/doc/html/draft-ietf-httpbis-priority-11#section-13.1> and Section 13.2 <https://datatracker.ietf.org/doc/html/draft-ietf-httpbis-priority-11#section-13.2> discuss mitigations
>    against this expansion of unfairness.
>
>
> There seems to be an (incorrect) implication that different progress
> between end-clients is unfair. However, given end-clients have different
> available bandwidths in the general case, it will be 'fair' that their
> rates are unequal. For instance:
> * Client A has a 100Mb/s downstream access bottleneck, but it is currently
> shared by 4 other applications
> * Client B is using a 120Mb/s downstream bottleneck exclusively for itself.
> It would be wrong here for an http intermediary to allocate equal
> bandwidth between two http applications run by these two clients, which
> would merely serve to take bandwidth from the other 4 applications, and
> waste some of B's capacity.
>
> This is the sense in which it's best to allow the transport layer to share
> out the bandwidth (which includes some ideas that never took off, like
> bandwidth brokers between applications). But the app layer is still
> controlling the bigger levers of bandwidth allocation.


[LP] Thanks. There's some good knowledge here that I'll defer to Kazuho to
respond on better than I. I'm open to nailing down definitions better if it
will help.

FWIW we give recommendations for servers to avoid starving on situations
like 5 non-incremental requests. Really, servers want to keep transport
connection bandwidth capacity as full as possible. The available capacity
is a product of various factors - HTTP/2 flow control can add a whole
dimension of implementer fun. If we imagine a scenario where the client
makes 5 non-incremental requests where the contents are written to 5
different disk drives; if response A gets flow-control blocked by the
client because the disk is full, then the server can switch to B if it
likes.


Cheers
Lucas


Bob

[1] - https://datatracker.ietf.org/doc/html/rfc7540#section-5.3.5