Santiago,
sorry for not being responsive for weeks, but we have completely renewed our
infrastructure in the past weeks which bound all my time. So here is my
proposal for improvements in Hypermedia Support. It is basing on yours found
at
http://java.net/projects/jax-rs-spec/pages/HypermediaExample, so I will
not describe the business scenario again but just post a proposed
alternative representation of the links.
Proposed Change #1: Representations and Links
In contrast to the referenced hypermedia example, this proposal omits the
use of a template but instead lists ready-to-use links for each machine. The
benefit is that a client can use this links even if the actual entity's
syntax would be unknown (i. e. the client cannot parse the machine IDs from
the entitiy) or would not contain the list of machines at all (e. g. if it
is a video stream showing the actual server rack, instead of a JSON
description of the rack's state). Also the client is not forced to do any
additional work like parsing and applying templates, but just can follow a
static link. The original example is unable to deal with this problems. As
JAX-RS should be as flexible as http itsels is, and as binary media more and
more is used (like videos and audios), JAX-RS should be able to work without
content parsing.
Content-Type:application/json
Link: </cluster>; produces="application/json"; method="GET"; rel="self"
</cluster/offliner>; produces="application/json"; method="POST";
rel="offliner"
</cluster/machine/alpha>; produces="application/json"; method="GET";
rel="item"
</cluster/machine/beta>; produces="application/json"; method="GET";
rel="item"
</cluster/machine/gamma>; produces="application/json"; method="GET";
rel="item"
{
"machines": [
{
"load": "1.4",
"name": "alpha",
"state": "STOPPED",
"nOfCpus": "2"
}, ...
],
"name": "cluster1",
"status": "ONLINE"
}
Actually I think it is not an either-or decision, but I propose that the
rendering of the links is by default non-templated, but can be optionally
switched to templated rendering using an annotation.
Please note that this representation is not my final proposal, but the
actual rendering should use the "anchor" attribute instead. The explanation
is dependent of the next change, so please read on to understand how the
idea develops!
Proposed Change #2: Separating navigation-target from transition-target
In analogy to a human-readable web site that shows a list of machines's
states and one pair of "startup" / "shutdown" buttons besides each machine,
one might want to contain such "non-self" transition links in the cluster's
representation (not in the machine's representation) to separate the
navigation-target (= next entity to get) from the transition-target (=
transitioned entity). (Side note: As a nice side effect, when using XML
instead of JSON this would apparently allow to simply create the fully
functional human-readable HTML page by just applying some XSLT ontop of our
JAX-RS service, BTW).
Content-Type:application/json
Link: </cluster>; produces="application/json"; method="GET"; rel="self"
</cluster/offliner>; produces="application/json"; method="POST";
rel="offliner"
</cluster/machine/alpha>; produces="application/json"; method="GET";
rel="item"
</cluster>; anchor="machine/alpha/stopper"; produces="application/json";
method="POST"; rel="stopper"
</cluster/machine/beta>; produces="application/json"; method="GET";
rel="item"
</cluster>; anchor="machine/beta/starter"; produces="application/json";
method="POST"; rel="starter"
</cluster/machine/gamma>; produces="application/json"; method="GET";
rel="item"
</cluster>; anchor="machine/gamma/stopper"; produces="application/json";
method="POST"; rel="stopper"
The above lines represent a cluster that has three machines: Alpha and gamma
are online, so there is only a offline contained. Beta is offline, so an
onliner is contained. In fact, we could just add more rel's for each machine
to allow more actions to be done on the "front page" of our cluster: like
"restart machine" as shown below:
Content-Type:application/json
Link: </cluster>; produces="application/json"; method="GET"; rel="self"
</cluster/offliner>; produces="application/json"; method="POST";
rel="offliner"
</cluster/machine/alpha>; produces="application/json"; method="GET";
rel="item"
</cluster>; anchor="machine/alpha/stopper"; produces="application/json";
method="POST"; rel="stopper"
</cluster>; anchor="machine/alpha/restarter"; produces="application/json";
method="POST"; rel="restarter"
</cluster/machine/beta>; produces="application/json"; method="GET";
rel="item"
</cluster>; anchor="machine/beta/starter"; produces="application/json";
method="POST"; rel="starter"
</cluster/machine/gamma>; produces="application/json"; method="GET";
rel="item"
</cluster>; anchor="machine/gamma/stopper"; produces="application/json";
method="POST"; rel="stopper"
</cluster>; anchor="machine/gamma/restarter"; produces="application/json";
method="POST"; rel="restarter"
Alpha and gamma are running, so can be restarted. Beta is down, so there is
still only the starter.
Following such a "non-self" transition link would startup / shutdown one
single machine but produce the cluster (!) view again (not navigating to
that machine). To reach the separation of transitioned object and navigation
target, the above lines show that the transitioned object is rendered using
the "anchor" as attribute described by RFC 5988. So, the invoked URI still
is naming the cluster's URI, not the machine's.
The last example shows that actually we could *always* use the anchor, even
with no additional transitions methods. Even if we just want to describe the
list of machines in the cluster:
Content-Type:application/json
Link: </cluster>; produces="application/json"; method="GET"; rel="self"
</cluster/offliner>; produces="application/json"; method="POST";
rel="offliner"
</cluster>; anchor="machine/alpha"; produces="application/json";
method="GET"; rel="item"
</cluster>; anchor="machine/alpha/stopper"; produces="application/json";
method="POST"; rel="stopper"
</cluster>; anchor="machine/alpha/restarter"; produces="application/json";
method="POST"; rel="restarter"
</cluster>; anchor="machine/beta"; produces="application/json";
method="GET"; rel="item"
</cluster>; anchor="machine/beta/starter"; produces="application/json";
method="POST"; rel="starter"
</cluster>; anchor="machine/gamma"; produces="application/json";
method="GET"; rel="item"
</cluster>; anchor="machine/gamma/stopper"; produces="application/json";
method="POST"; rel="stopper"
</cluster>; anchor="machine/gamma/restarter"; produces="application/json";
method="POST"; rel="restarter"
While it enlarges the header by a few bytes, it expresses more clearly the
fact that there is not necesserily any difference in rendering transitions
from navigations as Link headers, so the client does not need to handle them
differently in any way -- just as a browser being unaware of the fact
whether a button will transition state or navigate to a different resource
(or both). The sole logic (whether it is a transition or a navigation) is
completely under control of the server, so the client can be unaware of the
application's business logic (or even rather generic like a browser).
In fact, these lines above are my actual proposal how to render item lists
(see change #1)!
Regards
Markus
From: Santiago Pericas-Geertsen [mailto:Santiago.PericasGeertsen_at_oracle.com]
Sent: Mittwoch, 16. November 2011 00:22
To: jsr339-experts_at_jax-rs-spec.java.net
Subject: [jsr339-experts] Re: [jax-rs-spec users] Re: Re: Re: Re: Re:
Improving Hypermedia Support
Markus,
Inlined .
your proposal expects the client software to know the literals "onliner" and
"offliner" and their actual meaning. No machine can ever understand that
this will take machines offline. So at least these two words and their
actual intention must be known to the programmer of the client.
Loose coupling does not mean no coupling. All solutions would have some
degree of coupling. However, a solution in which the client needs to
understand the structure of all the links out-of-band simply does not
satisfy the most basic requirement of hypermedia IMO.
Also, he must know definitively how to apply the "item" template. As he must
know that, I cannot see what is more "loosely" and more "hypermedia" than
just sending the list of links directly? Just do not see the difference here
wrt to "hypermedia" and "coupling". So I do not agree to your point here.
What your proposal does better than mine is supporting "bad" design, as I
described later. With my proposal, there is less chance to do "bad" design.
And I think an API should prevent "bad" design if any possible.
Unless you show me a completely fleshed out example (like the one in the
wiki) using "your proposal", I simply cannot comment on whether mine is bad
and yours is good.
Please also not that my prosposal was not (a) and (b) as you cited, but
using the "anchor" attribute, which, BTW, I actually understand as the
intended solution to exactly this problem (separating target and method
without the need to invent another means of tempting).
As I said above, show us a complete example using "anchors" including the
Client code and I (and perhaps others on this alias) will be happy to
evaluate it.
-- Santiago
From: Santiago Pericas-Geertsen [mailto:Santiago.PericasGeertsen_at_oracle.com]
Sent: Montag, 14. November 2011 16:06
To: jsr339-experts_at_jax-rs-spec.java.net
Subject: [jsr339-experts] Re: [jax-rs-spec users] Re: Re: Re: Re: Improving
Hypermedia Support
On Nov 12, 2011, at 9:53 AM, Markus KARG wrote:
But, as a machine can neither guess nor understand tooltips, it just could
pick one of two strategies: (a) It is explicitly written to orchestrate this
particular service, so it just knows the syntax of the URI (I mean, if I
write an eBay client I obviously know the API, which contains the particular
URI syntax).
This strategy seems to be against hypermedia and loose coupling. Ideally,
we want clients to only know a single URI and then be dynamically informed
of the rest (as these may change even from run to run). I'm not saying that
URI templates is the best solution, but it's certainly better than clients
needing to know the exact URI structure to follow a link. Templates only
assume that clients know how to identify a machine and can call a method to
instantiate the template.
One doesn't need a template then (you only need a template, if you *do not*
know the syntax and want some background magic to handle this).
That's precisely the point, not knowing the syntax. If I know the exact
syntax, then there's no point of including these links in the
representations as they are already known to the client.
(a) Polymorhy: Maybe there are "machine" instances that need to have
differnent URIs! You would need to provide a different template, but how to?
(b) Binary content: Maybe the machine "name" is not (or not easily)
parseable from the content! How would your client software guess the machine
name to pass into the template?
Agree, it isn't perfect. But I can't agree that the strategy you outline
above is better.
But, actually, if I were you, I would provide a different model in fact to
solve your task: Not sending all the transition links for all the machines
inside of the cluster document, but let the cluster document contain only
the machines "self" URIs
This is certainly an option. In fact, given that what has been proposed are
all API calls, you could design your application this way --this is the
reason why I don't want to "bake" anything in annotations and framework
logic.
-- Santiago
(so you only learn the list of machines in the first step -- maybe you
neither want to know the state nor want to shut down a machine, so you don't
really lose anything in the first step). If a link is needed for a machine
(since you then want to know the status or want to shutdown one machine), I
can request the machine's representation in a second roundtrip and find it's
transition links inside that document. This imposes one more http request,
but it provides self-controlled machines, which in fact solves not only the
templates problem, the polymorphy problem and the binary content problem,
but it is also more scalable as for lots of machines you would only ask for
as much state as you can see on the screen (not requesting machines states
you won't read ever), and it is much more object oriented: If I want to shut
down a machine, I press it's power switch. I do not press a swith on a
different device. :-)
Regards
Markus
From: Santiago Pericas-Geertsen [mailto:Santiago.PericasGeertsen_at_oracle.com]
Sent: Dienstag, 8. November 2011 20:09
To: jsr339-experts_at_jax-rs-spec.java.net
Subject: [jsr339-experts] Re: [jax-rs-spec users] Re: Re: Re: Re: Improving
Hypermedia Support
On Nov 8, 2011, at 12:39 PM, Markus KARG wrote:
Can you please post the WIKI link again? Thanks.
Here it is:
http://java.net/projects/jax-rs-spec/pages/HypermediaExample
-- Santiago
From: Santiago Pericas-Geertsen [mailto:Santiago.PericasGeertsen_at_oracle.com]
Sent: Montag, 7. November 2011 19:11
To: jsr339-experts_at_jax-rs-spec.java.net
Subject: [jsr339-experts] Re: [jax-rs-spec users] Re: Re: Re: Improving
Hypermedia Support
Hi Markus,
Could you apply this to the example we have on the wiki? In particular,
given a collection item Y, how can I get the link to Y? I may want to start
machine A and stop machine B; so getting the links to all the machines in
the cluster is not enough.
-- Santiago
On Nov 6, 2011, at 8:45 AM, Markus KARG wrote:
Santiago,
this way:
Link: <absolute-uri-A1>; rel="name-of-collection-A"
Link: <absolute-uri-A2>; rel="name-of-collection-A"
Link: <absolute-uri-B1>; rel="name-of-collection-B"
Link: <absolute-uri-B2>; rel="name-of-collection-B"
Here you have four absolute URIs. A1 and A2 are part of collection A, B1 and
B2 are part of collection B.
If the client wants to visit all resources references by collection B, it
would just have to invoke A1 and A2, identified via collection A.
An alternative way would be using the anchor attribute to identify the
resource as a fragment of the source resource (e. g.
anchor="#name-of-collection-A").
A third option would be a link-extension like "Link: <absolute-uri>;
collection="name-of-collection".
Where do you see a problem? No need for neither templates nor client side
application logic.
Regards
Markus
From: Santiago Pericas-Geertsen [mailto:Santiago.PericasGeertsen_at_oracle.com]
Sent: Dienstag, 1. November 2011 20:17
To: jsr339-experts_at_jax-rs-spec.java.net
Subject: [jsr339-experts] Re: [jax-rs-spec users] Re: Re: Improving
Hypermedia Support
* I do not understand the problem with the collection. Why not sending just
several URIs with the same relation in that case (one for each item of the
collection)?
As link headers? How would you know which one corresponds to which item in
the collection? There's no ordering for link headers.
-- Santiago