Re: PRIORITY stream error?

[Ian Swett <[e-mail]>]

I think Roberto's principles make sense.

In this case, those principles imply the connection should be closed.  I
believe that's true of many of the HTTP/3 errors I'll note.

On Wed, Mar 6, 2019 at 6:40 PM Roberto Peon <[e-mail]> wrote:

> As a general principal (sometimes violated):
> If the behavior can be deterministically known to be incorrect, it should
> probably be a connection error.
>
> In cases where the behavior cannot deterministically be known to be
> correct (e.g. because of races/timing), it should not be a connection error.
>
> It seems like being "brutal" when possible (as per above) achieves the
> best interop.
> -=R
>
> On 3/6/19, 2:44 PM, "QUIC on behalf of Martin Thomson" <
> [e-mail] on behalf of [e-mail]> wrote:
>
>     This view (that the error is localized) is one that we took in
> HTTP/2.  In retrospect, I don't think that it has been used as much as you
> would think.  It is far easier to treat violations of spec brutally.
> Indeed, this tends to make the problem more visible, which is a good thing.
>
>     On Thu, Mar 7, 2019, at 09:37, Kazuho Oku wrote:
>     > 2019年3月7日(木) 5:59 Martin Duke <[e-mail]>:
>     > >
>     > > the very end of Section 4.2.3 of quic-http says:
>     > >
>     > >    PRIORITY frames received by a client MUST be treated as a stream
>     > >    error of type HTTP_UNEXPECTED_FRAME.
>     > >
>     > >
>     > > Elsewhere, this kind of thing is a connection error.
>     >
>     > I am not sure if I agree with the observation. IIUC, the general
>     > approach is to use stream errors when the error does not affect the
>     > entire connection.
>     >
>     > It is reasonable for a client to respond with a stream error when it
>     > observes a PRIORITY frame on a *request* stream.
>     >
>     > That said, I agree that it should be a connection error when the
>     > client receives a PRIORITY frame on a control frame. That's because
> we
>     > cannot have a stream-level error for a control stream, because the
>     > stream can never be closed. I think that's what is missing in the
>     > text.
>     >
>     > FWIW, we do have this "if the error is X then it's a stream-level
>     > error, or if the error is Y then it's a connection-level error" type
>     > of handling. See section 3.2.2 for an example.
>     >
>     > > Making this a  stream error seems problematic; if otherwise valid,
> if this goes out on the control stream a stream error may bring everything
> down anyway?
>     > >
>     > > Should this be a connection error, or am I missing something?
>     >
>     >
>     >
>     > --
>     > Kazuho Oku
>     >
>     >
>
>
>
>