If we provide an API like setTrailer(String name, Supplier<String>), then
the container can generate a trailers header that is sufficient for any
trailers that are eventually supplied, as there is no other mechanism
available to set trailers. If the app has also set a Trailers header,
then it can either be ignored or replaced. I prefer ignored as there are
many bad combinations of headers that an app can already generate and we
don't enforce RFC compliance for other headers.
If we provide and API like setTrailer(Supplier<Map<String,String>), then we
could look at any Trailers header set and only query the Map for listed
trailers. In that case we would probably be better off with
setTrailer(Function<String,String>) and avoid the map.
The requirement regarding the TE header is very subjective and hard for us
to enforce. Is a checksum "necessary for the user agent to receive" ?
Depends on the user-agent.
So options are:
a. setTrailer(Supplier<Map<String,String>) iterate over app set Trailer
headers
b. setTrailer(Function<String,String>), iterate over app set Trailers header
c. setTrailer(String name, Supplier<String>), ignore app set Trailers header
d. setTrailer(String name, Supplier<String>), clear app set Trailers header
e. setTrailer(Supplier<Map<String,String>) ignore app set Trailers header
(status quo)
My preferences are b,c,d,e,a
cheers
On 4 May 2017 at 02:01, Shing Wai Chan <shing.wai.chan_at_oracle.com> wrote:
> Let us summarize what we discuss on the trailer implementation.
>
> MT> The first one is triggered by Section 4.4 for RFC 7230:
>
> MT> <quote
> MT> When a message includes a message body encoded with the chunked
> MT> transfer coding and the sender desires to send metadata in the form
> MT> of trailer fields at the end of the message, the sender SHOULD
> MT> generate a Trailer header field before the message body to indicate
> MT> which fields will be present in the trailers. This allows the
> MT> recipient to prepare for receipt of that metadata before it starts
> MT> processing the body, which is useful if the message is being streamed
> MT> and the recipient wishes to confirm an integrity check on the fly.
> MT> </quote>
>
> MT> 1. How is the container meant to construct the Trailer header field?
> The
> MT> point of using a supplier was so the header names and values did not
> MT> have to be known when the response was committed. However, section 4.4
> MT> requires that the names are known.
>
> GW> Firstly, it's only a SHOULD, so I'm guessing it wont be well respected.
> GW> For example are we going to reject trailers not listed in a prior
> Trailers header?
> GW> But then I guess at the time the supplier is set, the app knows it is
> adding
> GW> a supplier with certain capabilities, so it should know what trailers
> may be set.
>
> MM> I think we should enforce the "SHOULD" at the servlet level so
> consider it
> MM> a must.
> MM> Also what if the application sets the trailer header using the
> existing
> MM> setHeader() api and then subsequently uses the new api to set the
> promised
> MM> trailer headers. Should we detect such a setting of the trailer header
> and
> MM> then ignore headers set using setTrailers if they were not indicated
> by
> MM> the trailer header.
>
> MT> Do we ignore the 'SHOULD' or do we change the API? I'm leaning
> MT> towards changing the API on the basis that if we are going to add
> MT> trailer support we should do it as correctly as possible.
>
> GW> +0 as it feels a bit over complicated. i guess we could do
> GW> setTrailer(String name, Supplier<String>)
> GW> To set a supplier for each field and then we'd know how to set the
> GW> Trailers header.
> GW> But to be symmetric, would we'd need to police the Trailers header on
> incoming
> GW> requests? symmetry also pushes us towards
> GW> String getTrailer(String name)
> GW> rather than the map.
>
> MM> As I said before I think we enforce the SHOULD. However I don't see
> MM> a need to be symmetric, we have no control over what clients send in
> MM> and we do the best we can with what we get. If the app really wants
> MM> the header field it can ignore any trailer headers but we should still
> MM> be willing to provide them.
>
> +1 for no need for symmetric for #getTrailerFields.
>
> In the previous API, Supplier<Map<String, String>>, it is the developer
> responsbility
> to generate a "trailer" header.
>
> In the above, setTrailerField(String name, Supplier<String> supplier), a
> "trailer"
> header will be generated by the servlet container.
>
> If we want to enforce the SHOULD at the servlet level, then we would like
> to consider
> the issue mentioned by Martin above:
> req.setHeader("trailers", "foo");
> req.setHeader("foo", "v1")
> // set a supplier generated trailer header "foo" with value "v2"
> The following are possible options:
> a) a header "foo: v1" is sent and a trailer header "foo: v2" is sent
> b) only a trailer header "foo: v1" is sent, ignore the trailer supplier
> c) only a trailer header "foo: v2" is sent, ignore the header
> d) throw IllegalStateException
>
> MM> I should also have mentioned this from the RFC section 4.1.2:
> MM> Unless the request includes a TE header field indicating "trailers"
> MM> is acceptable, as described in Section 4.3, a server SHOULD NOT
> MM> generate trailer fields that it believes are necessary for the user
> MM> agent to receive. Without a TE containing "trailers", the server
> MM> ought to assume that the trailer fields might be silently discarded
> MM> along the path to the user agent. This requirement allows
> MM> intermediaries to forward a de-chunked message to an HTTP/1.0
> MM> recipient without buffering the entire response.
> MM> Should we be looking for this header and throw an IllegalStateException
> MM> from setTrailers() if the TE header is not present or does not
> indicate that
> MM> trailers are acceptable?
>
> The possible options:
> i) throw IllegalStateException
> ii) ignore
>
> Let us discuss more on (1).
>
> ---
> MT> 2. What happens if setTrailerFields() is called after the response is
> MT> committed?
>
> GW> Good question. We can go either: If when it is set, the response is
> committed
> GW> with a transport mode that cannot support trailers then an
> illegalStateException
> GW> can be thrown; alternately we can just never call the supplier to get
> the trailers.
> GW> Jetty currently does the later, but I can see reasons for the former.
>
> MM> It is just the trailer header which must be sent before the response
> is
> MM> committed so in theory we could accept this after commit if the
> MM> trailer header was provided and sent before commit. This complicates
> things
> MM> but does not seem a unreasonable expectation from an app.
>
> MT> Throw an IllegalStateException
> MT> But. It also depends on 1). If we don't have to set the Trailers
> MT> header it may still be possible to send the trailer fields
>
> GW> +1
>
> +1
>
> ---
> MT> 3. What happens if setTrailerFields() is called and trailer fields are
> MT> not supported (HTTP 0.9, HTTP 1.0, some proxy protocols, etc.)?
>
> GW> Same as above.
>
> MM> Difficulty is if the trailer fields include some valid headers and
> some
> MM> invalid headers so I vote for just ignoring the invalid ones.
>
> MT> Throw an IllegalStateException
>
> GW> +1
>
> +1
>
> ---
> MT> 4. What happens if setTrailerFields() is called multiple times?
>
> GW> In Jetty, each call replaces the previous. So you can null out the
> supplier
> GW> or give a new one.
> GW> However, I can see that there might be competing concerns that want to
> add
> GW> multiple trailers. Perhaps we need addTrailerFields and support
> multiple
> GW> suppliers (a bit over the top), or perhaps just ISE if it has already
> been set?
>
> MM> I would go with replacing the previous, also if we replace a previous
> this
> MM> could definitely be done after commit, just to replace previous values.
>
> MT> Use the most recent Supplier
>
> GW> +1
>
> +1
>
> ---
> MT> 5. What happens if setTrailerFields() is called with a null Supplier?
>
> GW> Jetty currently nulls out the supplier.
>
> MM> +1
>
> MT> Allow it and protect against any possible NPE in the container
>
> GW> +1
>
> +1
>
> ---
> MT> Do we need a boolean areTrailerFieldsSupported() method?
>
> Not at this point unless (1) needs this.
>
> Thanks.
> Shing Wai Chan
>
>
>
> On May 3, 2017, at 2:17 PM, Martin Mulholland <mmulholl_at_us.ibm.com> wrote:
>
> I should also have mentioned this from the RFC section 4.1.2:
>
> Unless the request includes a TE header field indicating "trailers"
> is acceptable, as described in Section 4.3, a server SHOULD NOT
> generate trailer fields that it believes are necessary for the user
> agent to receive. Without a TE containing "trailers", the server
> ought to assume that the trailer fields might be silently discarded
> along the path to the user agent. This requirement allows
> intermediaries to forward a de-chunked message to an HTTP/1.0
> recipient without buffering the entire response.
>
> Should we be looking for this header and throw an IllegalStateException
> from setTrailers() if the TE header is not present or does not indicate
> that
> trailers are acceptable?
>
>
> Thank you,
> Martin Mulholland.
> WebSphere Application Server Web Tier Architect
> email: mmulholl_at_us.ibm.com
>
> IBM RTP, PO BOX 12195, 503/C227,
> 3039 Cornwallis Rd, RTP, NC 27709-2195
> t/l 444-4319, external (919)-254-4319
>
>
> Martin Mulholland/Raleigh/IBM_at_IBMUS wrote on 05/03/2017 05:04:04 PM:
>
> > From: Martin Mulholland/Raleigh/IBM_at_IBMUS
> > To: jsr369-experts_at_servlet-spec.java.net
> > Date: 05/03/2017 05:05 PM
> > Subject: [servlet-spec users] [jsr369-experts] Re: Re: Trailer
> > header implementation
> >
> >
> > Thank you,
> > Martin Mulholland.
> > WebSphere Application Server Web Tier Architect
> > email: mmulholl_at_us.ibm.com
> >
> > IBM RTP, PO BOX 12195, 503/C227,
> > 3039 Cornwallis Rd, RTP, NC 27709-2195
> > t/l 444-4319, external (919)-254-4319
> >
> >
> > Greg Wilkins <gregw_at_webtide.com> wrote on 05/03/2017 08:54:07 AM:
> >
> > > From: Greg Wilkins <gregw_at_webtide.com>
> > > To: jsr369-experts_at_servlet-spec.java.net
> > > Date: 05/03/2017 08:54 AM
> > > Subject: [servlet-spec users] [jsr369-experts] Re: Trailer header
> > > implementation
> > >
> > >> On 3 May 2017 at 13:02, Mark Thomas <markt_at_apache.org> wrote:
> > >> Hi,
> > >>
> > >> I've been working on trailer header implementation. No particular
> issues
> > >> with requests but the response implementation is identifying a lot of
> > >> questions.
> > >>
> > >> The first one is triggered by Section 4.4 for RFC 7230:
> > >>
> > >> <quote
> > >> When a message includes a message body encoded with the chunked
> > >> transfer coding and the sender desires to send metadata in the form
> > >> of trailer fields at the end of the message, the sender SHOULD
> > >> generate a Trailer header field before the message body to indicate
> > >> which fields will be present in the trailers. This allows the
> > >> recipient to prepare for receipt of that metadata before it starts
> > >> processing the body, which is useful if the message is being
> streamed
> > >> and the recipient wishes to confirm an integrity check on the fly.
> > >> </quote>
> > >>
> > >> 1. How is the container meant to construct the Trailer header field?
> The
> > >> point of using a supplier was so the header names and values did not
> > >> have to be known when the response was committed. However, section 4.4
> > >> requires that the names are known.
> >
> > >
> > > Firstly, it's only a SHOULD, so I'm guessing it wont be well
> > > respected. For example are we going to reject trailers not listed
> > > in a prior Trailers header?
> > >
> > > But then I guess at the time the supplier is set, the app knows it
> > > is adding a supplier with certain capabilities, so it should know
> > > what trailers may be set.
> > >
> >
> > I think we should enforce the "SHOULD" at the servlet level so consider
> it
> > a must.
> >
> > Also what if the application sets the trailer header using the existing
> > setHeader() api and then subsequently uses the new api to set the
> promised
> > trailer headers. Should we detect such a setting of the trailer header
> and
> > then ignore headers set using setTrailers if they were not indicated by
> > the trailer header.
> >
> > >
> > >> The remaining questions are a little simpler:
> > >>
> > >> 2. What happens if setTrailerFields() is called after the response is
> > >> committed?
> > >
> > > Good question. We can go either: If when it is set, the response
> > > is committed with a transport mode that cannot support trailers then
> > > an illegalStateException can be thrown; alternately we can just
> > > never call the supplier to get the trailers.
> > >
> > > Jetty currently does the later, but I can see reasons for the former.
> > >
> >
> > It is just the trailer header which must be sent before the response is
> > committed so in theory we could accept this after commit if the
> > trailer header was providedand sent before commit. This complicates
> things
> > but does not seem a unreasonable expectation from an app.
> >
> > >
> > >> 3. What happens if setTrailerFields() is called and trailer fields are
> > >> not supported (HTTP 0.9, HTTP 1.0, some proxy protocols, etc.)?
> > >
> > > Same as above.
> >
> > Difficulty is if the trailer fields include some valid headers and some
> > invalid headers so I vote for just ignoring the invalid ones.
> >
> > >
> > >> 4. What happens if setTrailerFields() is called multiple times?
> > >
> > > In Jetty, each call replaces the previous. So you can null out the
> > > supplier or give a new one.
> > >
> > > However, I can see that there might be competing concerns that want
> > > to add multiple trailers. Perhaps we need addTrailerFields and
> > > support multiple suppliers (a bit over the top), or perhaps just ISE
> > > if it has already been set?
> > >
> >
> > I would go with replacing the previous, also if we replace a previous
> this
> > could definitely be done after commit, just to replace previous values.
> >
> > >
> > >> 5. What happens if setTrailerFields() is called with a null Supplier?
> >
> > > Jetty currently nulls out the supplier.
> > >
> >
> > +1
> >
> > >
> > >> My current thinking:
> > >>
> > >> 1) Do we ignore the 'SHOULD' or do we change the API? I'm leaning
> > >> towards changing the API on the basis that if we are going to add
> > >> trailer support we should do it as correctly as possible.
> > >
> > > +0 as it feels a bit over complicated. i guess we could do
> > >
> > > setTrailer(String name, Supplier<String>)
> > >
> > > To set a supplier for each field and then we'd know how to set the
> > > Trailers header.
> > >
> > > But to be symmetric, would we'd need to police the Trailers header
> > > on incoming requests? symmetry also pushes us towards
> > >
> > > String getTrailer(String name)
> > >
> > > rather than the map.
> >
> > As I said before I think we enforce the SHOULD. However I don't see
> > a need to be symmetric, we have no control over what clients send in
> > and we do the best we can with what we get. If the app really wants
> > the header field it can ignore any trailer headers but we should still
> > be willing to provide them.
> >
> > >
> > >> 2) Throw an IllegalStateException
> > >> But. It also depends on 1). If we don't have to set the Trailers
> > >> header it may still be possible to send the trailer fields
> > >
> > > +1
> > >
> > >
> > >> 3) Throw an IllegalStateException
> > >
> > > +1
> > >
> > >> 4) Use the most recent Supplier
> > >
> > > +1
> > >
> > >> 5) Allow it and protect against any possible NPE in the container
> > >
> > > +1
> > >
> > >> Of these, 1) is the most serious to address. After that, 2) & 3) since
> > >> the application has no way to tell if this is going to happen. Do we
> > >> need a boolean areTrailerFieldsSupported() method?
> > >>
> > >> Mark
> > >
> >
> > >
> > > --
> > > Greg Wilkins <gregw@webtide.com> CTO http://webtide.com
>
>
>
--
Greg Wilkins <gregw@webtide.com> CTO http://webtide.com