Quoting [Sam Johnston] (May 13 2009):
>
>> Yes, me, I don't think HATEOAS should be applied in this
>> context. But I realise/accept that I maybe the only one
>> with that opinion - thats ok. So I'll say it here one last
>> time, for the record, and then will shut up: "a static
>> simple state model allows for very simple clients.
>> Extensions can be defined via substates, or additional
>> transitions."
>
> I would counterargue that HATEOAS allows for even simpler clients
> because they don't have to worry about hardwiring even a simple state
> model. Using HTTP we can even feed them plain $LANG descriptions of
> what the transitions and targets are - it doesn't get any easier than
> that and you don't have to worry about updating clients to implement
> new goodies.
I don't see that. If I want to write a client tool which
starts a resource, I want to make sure the resource is in
RUNNING state when the client reports success. But if that
client (a) has to infer the available states from a
registry, it cannot posisbly know which state has the
semantic meaning of RUNNING attached. Further (b), if the
client only sees those state transitions it is allowed in
its current state, how does it know what transition path to
take to reach that target state? Is it (I am making those
up obviously):
INITIAL -> create() -> CREATED -> elevate() -> ELEVATED () -> run() -> RUNNING
or
INITIAL -> create() -> CREATED -> init() -> INITIALIZED -> run() -> RUNNING
Or should the tool simply fail because it cannot see a run()
transition in its INITIAL state?
The client must at least know how to create a resource and when it has done so successfully a "start" actuator will appear, perhaps with a target state of "running" (TBD). In that case it knows that if it pulls the "start" handle eventually the resource should end up "running". Otherwise it could know (from the registry) that "start" is the right button to push, but that's starting to break HATEOAS principles. We have options - it's just a matter of finding the right one.
I think HATEOAS works pretty well if a human is in the loop
who can parse the available transition description, and
deduce a semantic meaning. I don't think it makes for
simple tooling, really.
I agree that humans are better at this stuff than computers but I'm unconvinced this translates to complex tooling.
Then again, I may misunderstand the proposed usage of
HATEOAS in OCCI. So, can you help me out: what mechanism
will avoid the confusion from the example above, if a vendor
can provide init() and elevate() transitions on the fly,
with no predefined semantics attached? How would my tool
deduce the transition path it needs to enact?
The semantics for common functions will be in the registry. It's ones that are uncommon and impossible to predict like "translate" and "migrate" that we're catering for here, and generally there will need to be some kind of client side support for these.
As I said below, "we may need to revisit this point in the name of interop", and I suggested categories as one possible solution (e.g. a "starting" vs a "stopping" transition)... parametrised transition calls are another... for example, how do I tell something to start *without* saved state if saved state is present (ala cold start vs resume)?
Sam
> I don't think anyone knows every possible thing that users are going to
> want to do with the API (I certainly don't have the confidence to say I
> do anyway) but we may need to revisit this point in the name of
> interop... Atom categories would be one way to achieve this (e.g. "Cold
> Reboot" and "Warm Reboot" might go in the "restart" category).
> Sam
>
> References
>
> 1. mailto:andre@merzky.net
--
Nothing is ever easy.