Re: Design Issue: Max Concurrent Streams Limit and Unidirectional Streams

[William Chan (陈智昌) <[e-mail]>]

I think there's some missing context not being stated in the emails which
is leading to confusion. Can we revisit why decoupling the stream
directions will lead to a longer header lifetime? I would think that when
the streaming layer encounters headers for a stream, it would hand it up to
the HTTP semantic layer. As long as the HTTP semantic layer understands the
request-response coupling, it knows to keep the request headers around
until the response arrives, right?


On Fri, May 3, 2013 at 10:49 PM, Roberto Peon <[e-mail]> wrote:

> Hey, you're the one worried about the size of the compressor state (which
> would be ~4 or 8k)! :)
> Headers are sometimes larger individually, and the upper limit to the size
> of this state is the sum of 10s to 100s of these.
>
> I think that, pragmatically, since there is a framing-layer solution which
> ensures that one must not store headers for longer that necessary, and
> which is semantic-layer agnostic, it is a decent bet.
> Any approach other than declaring it as unidirectional (or equivalent)
> either requires the caching of much state, or a NACK from the remote side.
>
> That isn't to say that Martin's approach doesn't have appeal. I want to
> like it, but unless we are to have unlimited streams in some limbo state,
> it would require a NACK, else I won't be able to correlate sets of headers
> on a stream and the NACK both requires more machinery at both ends, and
> consumes more bytes on the wire.
> -=R
>
>
>
> On Fri, May 3, 2013 at 6:19 PM, James M Snell <[e-mail]> wrote:
>
>> Ok.. going back over the thread in detail and over the spec again, one
>> approach to addressing the overall concern here (and hopefully bring a
>> bit more rigor to the overall design) is to redefine the stream states
>> slightly along the same lines already suggested by Martin. Each
>> endpoint would maintain its own view of the current activity state of
>> every stream in a session, however, the state would only reflect the
>> actions taken by the peer endpoint. There are five possible activity
>> states:
>>
>> Unused
>>   The peer endpoint has not reserved or used the stream in any way.
>>
>> Open
>>   The endpoint has received frames on the stream from the peer, none
>> of which are type RST_STREAM or have the FINAL flag set.
>>
>> Closed
>>   The endpoint has received an RST_STREAM frame, or any frame with the
>> FINAL flag set from the peer
>>
>> Reserved-Open
>>   The peer has reserved the stream identifier for future use but
>> frames have not yet been received on that stream. The receiving
>> endpoint is expected to send its own frames on the same stream.
>>
>> Reserved-Closed
>>   The peer has reserved the stream identifier for future use but
>> frames have not yet been received on that stream. The receiving
>> endpoint is not expected to send its own frames on the same stream.
>>
>> MAX_CONCURRENT_STREAMS == The number of streams in the Open state the
>> endpoint will permit the peer to initiate at any given time. Once that
>> limit is reached, the receiving endpoint will likely begin rejecting
>> new streams using RST_STREAM. In other words, right now,
>> MAX_CONCURRENT_STREAMS is defined in terms of what the sending
>> endpoint must not do. This changes the definition to an indication of
>> what the receiving endpoint will do once a particular threshold is
>> reached. Any endpoint that wants to be able to keep creating streams
>> must be diligent about sending FINAL frames, etc.
>>
>> As for the Request-Response bounding issue, that's really an HTTP
>> semantic layer notion. I'm not fully convinced we really need to
>> handle that issue in the framing layer at all.
>>
>>
>> On Fri, May 3, 2013 at 2:20 PM, Roberto Peon <[e-mail]> wrote:
>> > The biggest rub in Martin's suggestion is that, as a stream initiator,
>> I no
>> > longer know for how long I should keep the original "request" headers
>> > around.
>> > I view that as an annoying problem (I want every response to be
>> attributable
>> > to a request).
>> >
>> > I also think it is a bit confusing-- how would it be used in cases where
>> > I've sent all my data on what I thought was a unidirectional stream, and
>> > then receive bytes from the other side on that stream. That'd be...
>> weird.
>> >
>> > With the unidirectional bit (or similar declaration of half-closed
>> > start-state), I now know (by fiat, essentially) that I will not receive
>> a
>> > response on that stream ID, and so I don't need to keep the "request"
>> > headers around after I've finished pushing the stream. Logging
>> accomplished.
>> >
>> >
>> > I think this is an easy issue to solve by reinstating the
>> unidirectional bit
>> > (for now). It is certainly minimal work to have servers which do server
>> push
>> > set that bit.
>> >
>> > To Will's point, I agree that an "ENHANCE YOUR CALM" code seems
>> redundant.
>> > In my case I believe it redundant because the remote side has already
>> > received my settings frame, or is sending without having known it (i.e.
>> > within the initial RTT), and will be receiving the SETTINGS frame
>> before it
>> > could process this new code anyway (assuming I'm following spec and
>> sending
>> > SETTINGS immediately upon session establishment).
>> > -=R
>> >
>> >
>> >
>> >
>> >
>> > On Fri, May 3, 2013 at 11:28 AM, William Chan (陈智昌) <
>> [e-mail]>
>> > wrote:
>> >>
>> >> I guess I kinda think that we're worrying too much about this corner of
>> >> the spec. I don't view it as a big deal in practice. The problem
>> described
>> >> happens when MAX_CONCURRENT_STREAMS is too low to allow enough
>> parallelism
>> >> per roundtrip. I would advise people to simply increase their
>> >> MAX_CONCURRENT_STREAMS in that case. I kinda think this is only
>> problematic
>> >> when we have very high latencies and devices that can't handle high
>> >> parallelism, like an interplanetary refrigerator that speaks HTTP/2
>> for some
>> >> reason. <shrug>
>> >>
>> >> I am unsure how to feel about a ENHANCE YOUR CALM code as it's not well
>> >> defined. I don't mind RST_STREAMs on exceeding limits, like the initial
>> >> MAX_CONCURRENT_STREAMS, since they're usually the result of a race (the
>> >> possible initial SETTINGS frame race) and we won't have to keep
>> continually
>> >> sending RST_STREAMs to rate limit appropriately.
>> >>
>> >>
>> >> On Fri, May 3, 2013 at 3:02 PM, James M Snell <[e-mail]>
>> wrote:
>> >>>
>> >>> The impact on client-to-server initiated streams is another reason why
>> >>> I suggested the credit-based approach and why it would likely be good
>> >>> to have an RST_STREAM "ENHANCE YOUR CALM" error code [1]. If the
>> >>> client misbehaves and sends too much too quickly, we have flow
>> >>> control, settings, rst_stream and goaway options to deal with it.
>> >>>
>> >>> [1]
>> >>>
>> http://en.wikipedia.org/wiki/List_of_HTTP_status_codes#4xx_Server_Error
>> >>>
>> >>> On Fri, May 3, 2013 at 10:34 AM, William Chan (陈智昌)
>> >>> <[e-mail]> wrote:
>> >>> > As I understand the proposal, which I believe ties into the issue
>> James
>> >>> > raised at the beginning here, the goal is to be able to open and
>> close
>> >>> > a
>> >>> > directional stream without an ACK, which I am nervous about as I
>> said
>> >>> > above
>> >>> > without much detail. Concretely speaking, a HTTP GET is a
>> >>> > HEADERS+PRIORITY
>> >>> > frame with a FINAL flag or an extra DATA frame with FINAL flag. This
>> >>> > means
>> >>> > that the request effectively never gets counted against the
>> directional
>> >>> > stream limit, as controlled by the receiver which sends a
>> >>> > MAX_CONCURRENT_STREAMS setting, since it open and closes the
>> direction
>> >>> > in
>> >>> > the same frame (or closes in the subsequent empty DATA frame).
>> >>> >
>> >>> >
>> >>> > On Fri, May 3, 2013 at 1:52 PM, Martin Thomson
>> >>> > <[e-mail]>
>> >>> > wrote:
>> >>> >>
>> >>> >> On 3 May 2013 09:44, William Chan (陈智昌) <[e-mail]>
>> wrote:
>> >>> >> > I'd like server folks to chime in, but doing this makes me feel a
>> >>> >> > bit
>> >>> >> > nervous. I feel this effectively disables the directional
>> concurrent
>> >>> >> > streams
>> >>> >> > limit. The bidirectional full-close essentially acts like an
>> ACK, so
>> >>> >> > removing it might result in an unbounded number of streams.
>> >>> >>
>> >>> >> I think that I know what you mean here, but can you try to expand a
>> >>> >> little?  Do you refer to the possible gap between close on the
>> >>> >> initiating direction and the first frame on the responding
>> direction;
>> >>> >> a gap that might cause the stream to escape accounting?  I think
>> that
>> >>> >> is a tractable problem - any unbounded-ness is under the control of
>> >>> >> the initiating peer.
>> >>> >
>> >>> >
>> >>
>> >>
>> >
>>
>
>