On Mar 6, 2013, at 9:31 PM, Denis Gervalle <[email protected]> wrote:
> Hi Vincent,
>
>
> On Wed, Mar 6, 2013 at 4:48 PM, Vincent Massol <[email protected]> wrote:
>
>>
>> On Mar 6, 2013, at 1:02 PM, Denis Gervalle <[email protected]> wrote:
>>
>>> Hi Vincent,
>>>
>>> On Wed, Mar 6, 2013 at 11:36 AM, Vincent Massol <[email protected]>
>> wrote:
>>>
>>>> Resending since I've made mistakes (it's only about ObjectReference, not
>>>> Properties), sorry about that. Here's the new version:
>>>>
>>>> ------------
>>>>
>>>> Hi devs,
>>>>
>>>> ATM in the model module there's no ability to reference an xobject other
>>>> than by using a free form name.
>>>
>>>
>>> We made this on purpose.
>>
>> I know, I worked on this… but it doesn't make it right… :)
>>
>
> So from this and what I read below, I understand you have suddenly decide
> to revert the new model back to the old way of addressing objects and stop
> using named objects. This is a major change that should be discussed first
> and does not have my support right now.
No I haven't changed that. I'm proposing to have both ways. Either by name or
by class ref+index. Same as currently.
>>>> The problem is that this is not really usable. This is why we introduced
>>>> the BaseObjectReference in oldcore.
>>>>
>>>
>>> We have introduce the BaseObjectReference in oldcore simply to easily
>>> support the current storage implementation. At that time, we had a
>>> discussion about using the object number (not to be confuse with the
>> index)
>>> or the object UUID. And, since the object UUID was suspicious in some
>> edge
>>> case, you convince me to use the object number to create the free form
>> name
>>> of these old objects (I should have resist more since I still do not have
>>> evidence of these edge case).
>>
>> What? That defeats one of the main goal of references which is that you
>> can directly address an entity without having retrieved it first!
>>
>
> True for the old object storage, but was a better a way to avoid what you
> seems to advocate now !
>
>
>> The reason entities have names is to be able to easily reference them.
>
>
> Ah, you are right, object should have name to be addressable properly.
>
>
>> Using a UUID makes it impossible to reference something without getting it
>> first...
>>
>
> Was only about a temporary compatibility solution for old objects,
> obviously not a good way to be able to create reference from scratch, but a
> good way to get stable reference from objects.
>
>
>> If to reference an Object I have to load the Document first, then in
>> practice it means we don't need Object references and they shouldn't be
>> considered entities.
>
>
> I disagree, you do not always have to load the document to know an object
> UUID. You may retrieve the UUID of object by searching the database. This
> is the only way I have found until now to strongly link object together
> when multiple objects are stored in the same document. However, most of the
> time, I avoid this by creating only a single object of a given class per
> documents when I want to reference them elsewhere. So I may use the
> document itself for reference.
>
>
>> For any entity, you should be able to construct a direct reference to it
>> without needing to load anything else.
>>
>
> This is precisely why we need name for referencing our objects.
Either name or class ref + index.
In the majority of use cases people are not going to use names and we still
need to be able to create a reference to an object easily and class ref + index
is the best solution I know of for that.
>> In the 7-8 years of XWiki we've used Class reference and numbers/position
>> to reference xobjects and it has worked quite well IMO.
>
>
> I use it since as long as you, and I have mostly worked without any
> reference to objects since these were missing. I mean I have usually not
> accessed object unless I have them at hand, since the only correct way to
> get them was exactly what you explain just above. I do not remember having
> really retain object number to reference them at any moment except for a
> few line of code. I have surely never based anything on "this is the third
> object of that class in that document" since this is not really stable in
> the current model.
Yes the concept of ref did not exist but the old apis are doing the same, i.e.
finding objects by class ref + number.
>> In most cases you have only 1 xobject of a type and it's very nice to be
>> able to say: here's a reference to the first object of type N in document P.
>>
>
> I said that I agree with this special case, which is really different IMO
> than accessing the nth object of document. We had for this special case
> helper function in XWiki since a long time and these revealed to be helpful.
yup
>> Exposing an internal id of a document as reference is a no go for me since
>> it cannot be constructed without retrieving the entity first.
>>
>
> That is not my purpose, it was only my temporary solution waiting for truly
> named objects, it have been decided to use object numbers (not index !),
> but this has nothing to do with the current discussion anyway. Maybe this
> decision influence you however.
>
>
>>>> However this is major PITA since we can't have clean code that create an
>>>> object reference and that doesn't depend on oldcore.
>>>>
>>>
>>> Creating a reference to the third object of a given class in a document
>> (or
>>> object number 3, or even the third object of a document) has absolutely
>> no
>>> meaning unless you have already that object at hand
>>
>> Why doesn't it have a meaning? If in my app, my spec says:
>> - each doc has 3 objects of type P and the first one means this, the
>> second one that and the 3rd one that other thing,
>> then the 3rd one has a meaning.
>>
>
> And when any object get deleted, your meaning vanished. Great design of
> course !
errr? Whether it's the first one or the Nth, if it's deleted then it's no
longer there obviously :)
Same, if you delete a doc and there's a ref to it, you'll get a non existing
document… nothing special here...
>>> , and so you already
>>> have a source for its reference (solved not nicely by the
>>> BaseObjectReference actually, but this was another story).
>>>
>>>
>>>> I'd like to propose the following:
>>>> * Modify ObjectReference to add 2 named parameters: Class reference and
>>>> position
>>>>
>>>
>>> Since position is really fragile without any real meaning,
>>> having a
>>> reference using a position will be a source for spurious issues, I am
>>> definitely -1 for any positional reference.
>>>
>>>
>>>> * Make the name optional in EntityReference
>>>>
>>>
>>> This sounds like a sign of bad design to me... what is a entity without a
>>> name in general ?
>>
>> It means it's a reference to an entity by a mean other than a name :)
>>
>
> The mean is bad, the position may change each time object are added or
> deleted. And I do not think we have real positioning in the current model,
> which means you cannot really ensure object order at any time.
AFAIK we do have an order:
private Map<DocumentReference, List<BaseObject>> xObjects = new
TreeMap<DocumentReference, List<BaseObject>>();
This means that for each class ref, there's an ordered list.
>> We need to decide but we could very well decide that some entities don't
>> have a name and that they can be found/addressed by some other means (as in
>> the case of Object references when locating them through class reference
>> and position).
>>
>
> We may, but we need strong reason to introduce that IMO. And other means
> should be solid, position is not.
It's almost as "solid" (or as "weak") as names for documents. Those docs can be
renamed and your ref is gone!
As soon as you use names and not ids in references you have "weak" referencing,
be it for objects or any other entity. That's not a problem. It has the
advantages that refs can be constructed.
The only other solution I can think of (which wouldn't work with the current
model and thus we would need to modify the current model to make it work before
we can implement the new model api with the old model impl) would be to
**force** always having a name for Objects when they are created and when the
user doesn't specify the name it's computed automatically using a default
scheme. For example: <class ref wiki>:<class ref space>.<class ref
page>[number] where [number] isn't specified for the first xobject but only for
the next ones.
For example for a class ref of (wiki = my:wiki, space = my.space, page =
mypage) we would have "my:xwiki:my.space.mypage" for the first object and
"my:xwiki:my.space.mypage[1]" for the second one, etc.
Note that it wouldn't be a serialized ref since we need to make it as easy as
possible to create it using string concatenation. It should never be used as a
serialized class ref and not used to find the class ref. It's just a string.
Moving the object around would keep this name (unless the user manually changes
it). And renaming the xclass document would also not change this name.
So here's what we would need to write to get the avatar property for a user:
DocumentReference userDocumentReference = ...
ObjectReference userObjectReference = new ObjectReference("XWiki.XWikiUsers",
userDocumentReference);
ObjectPropertyReference avatarReference = new ObjectPropertyReference("avatar",
userObjectReference);
String fileName = this.entityManager.getEntity(avatarReference).getValue();
Again, if we want to go this way we need to modify the oldcore to add internal
support for names to XObjects. This means modifying the DB schema, not a very
small change...
Which is why I think that using class ref + index, *in addition to name* is
probably better for now.
WDYT?
see more below
>> More below
>>
>>>> This means that when we use an EntityReferenceResolver to resolve
>>>> "wiki:space.page^wiki2:space2.page2" we get an ObjetReference with:
>>>> * name = null
>>>> * param1: name = "classReference", value = EntityReference
>>>> * param2: name = "objectPosition", value = 0
>>>>
>>>
>>>> Rationale:
>>>> * This is exactly what we already do for Locale (and what we'll do for
>>>> Version too probably) so it's logical to do it for Object References too
>>>>
>>>
>>> I agree with your sample and your rationale, there is a need to create a
>>> reference to the first (and probably the only for such use case) object
>> of
>>> a given class in a given document without having to compose weird names
>> or
>>> positional references. This has definitely a meaning, much more than the
>>> second or any numbered objects... And this object is not always "[0]" in
>>> base reference syntax !
>>
>> ah good, was trying to despair :)
>>
>> If the first one has a meaning, then 2nd or 3rd can also have a meaning,
>> see my example above. Obviously the majority of use cases will be the first
>> one but that doesn't mean the other positions are not needed.
>>
>> 0 = first not null xobject
>> 1 = second not null xobject
>> ...
>>
>
> When you use the first one, and have only one, you may ensure you get the
> right one whatever the creation/deletion of object happened. This is not
> true for the other one, and we do not have anything for that.
> The reason I am -1 for using position, is mainly because position may
> change overtime for the same object, so this is a really fragile reference,
> and obviously this is not good design at all for maintaining a reference.
> Position is perfect during a small piece of code, and bad to be stored as a
> reference.
>
>
>>
>>> However, resolving "wiki:space.page^wiki2:space2.page2" to that object is
>>> not valid, since you do not really know what you are doing here, are you
>>> speaking about the first object of a given class or an object named "
>>> wiki2:space2.page2" ?
>>
>> Exactly. Which is why in my proposal we would have to agree that the
>> current notation (thomas calls it syntax) only describes class reference
>> and NOT a name (name = null). If we ever want to also be able to write a
>> string and specify a name in it then we would need to invent a syntax, in
>> the same manner that we currently have no syntax to express a locale in a
>> document reference when expressed as a string.
>>
>
> I see it the other way round.
>
>
>>> So if we need to resolve a string into the first (or,
>>> even if I am against, any) object of a given class, we need another
>> syntax
>>> at least (BaseObjectReference had already cause some poor stuff in
>>> LocalUidReferenceSerializer where we had to remove wiki in a very bad way
>>> currently).
>>
>> Using a name is very very far in the future so we don't need to invent a
>> syntax for that now IMO.
>
>
> If it is, what is the purpose of renewing the model ? Why do you prefer to
> step backward instead of going forward to solve our initial issue ?
It's not a step backward. We still have the name while at the same time making
it possible to also reference objects by a construction (class ref + index).
BTW I don't consider class ref + index a bad thing. IMO it was pretty clever.
Doing a new model doesn't mean throwing out all good ideas that existed in the
old model :)
>> We don't even know if it'll ever happen actually…
>>
>
> This really afraid me, no more ambition to get it right ?
No, you don't get it at all...
As a user I don't **care at all** about object names in 99% of the cases. I
don't want to name them as a user. That's a pain and a constraint uses
shouldn't care about.
Again we don't do that now and it works very well.
>>>> Consequences:
>>>> * We need to modify the Seralizers/Resolvers accordingly
>>>>
>>>
>>> According we have a new syntax for your kind of object reference, we may
>> do
>>> so. I resolver/serializer (as well as setter of class reference) I
>> suggest
>>> like Thomas that we support relative reference to classes based on the
>>> containing document, both simplifying creation and avoiding
>>> useless repetition.
>>
>> That seems good for now, even though in some far future we may want to
>> have a class defined in a wiki and an xobject located in another subwiki…
>> or not… ;)
>>
>
> Not repeating useless information does not mean we should not support
> absolute reference as well. Thinking about it twice, I agrre with you that
> this is more subtle than it look, since copying a document may be greatly
> influence by this. We should discuss that further once we agree on the rest.
>
>
>>
>>>> * We need to modify EntityReference to support a null name
>>>>
>>>
>>> Do we really need that, or the name would simply reflect our special
>> syntax
>>> ?
>>
>> It would be slightly misleading to use the serialized class ref as the
>> name IMO but why not... One danger is in code that will use it instead of
>> using the well-known parameters for class ref and position.
>>
>
> You cannot always prevent bad coding practice, and you may also think about
> the easy way some may want to use in velocity for example even if this is
> not optimal.
>
>
>>
>>>> * We deprecate BaseObjectReference
>>>>
>>>
>>> Does it really true ? It does not have the same meaning (even for your
>>> initial proposal) since number and position are not the same and the
>>> BaseObjectReference is a special reference, not a general object
>> reference,
>>> since it does not have a real free form name. Deprecating means we had to
>>> replace them everywhere, and I am not sure it is easy to use index in
>> place
>>> of number in some places, like storage where we use reference for IDs.
>>
>> Maybe, in any case what I meant is that we'll need to make our code use
>> the new way and not use BaseObjectReference as much as possible.
>>
>>>> * Probably some other stuff to modify like modifying event listeners
>>>> listening on objects since it's now going to be much easier, etc
>>>>
>>>> WDYT?
>>>>
>>>
>>> So, to resume, I understand the need to have a way to create reference to
>>> the first object of given class in a document. Even if using an object
>>> reference with, let say a special syntax, will introduce some bad
>>> consequences on the implementation of equals
>>
>> equals/hashcode shoudl already work since they already take into account
>> parameters.
>>
>
> But comparing two references to the same object, one using the name of the
> object (once we have it of course) and the other the classname will not
> compare to be equal while these are. We have the exact same situation for
> locale or version.
They won't be the same reference, which is ok. You can have several refs
leading to the same entity. There's no issue about that.
>>> and related methods on entity
>>> reference (like those we have for locale or version), we should probably
>>> consider it (Note that it will also affect property reference).
>>>
>>> I am definitely -1 for any positional reference which are meaningless.
>>
>> I think this is a pity and you're closing the door to lots of use cases
>> and you're going to make them really hard and not performant to implement
>> as a consequence.
>>
>
> I am not closing anything, I am just not agreeing to create a solution that
> will encourage bad design.
>
> As a user, if I **want** to access the 3rd xobject because it has a meaning
>> for my app, I'm not going to be able to do so.
>
>
> As a user I want to access the right object, I do not want to care about
> its place, I just need to be able to name it properly. Your use case is
> simply wrong and induced by the current design.
>
>
>> I'm going to need to load the doc and do some queries to iterate over
>> xobject and count to 3… (which obviously is both painful and not
>> performant).
>>
>
> And is need it anyway to access that object, it is you or the API, but it
> will be done at some point.
>
>
>> BTW your counter-suggestion to drop the position means you're going to
>> break everything that already exists since it's currently possible to
>> reference an object with class and number. It means inventing a new syntax,
>> which is always painful to do.
>>
>
> Have I said you cannot use the BaseObjectReference ?
> You may do so, but you are stick with the oldcore and the old and bad way
> of working with object.
> Again, why not using our energy to find a nice way to reach the named
> object solution instead of defending a bad design we suffer from since
> really too long ?
But it's not bad at all. It's very good! Forcing users to create a name for all
objects they create would be a very bad solution… Autogenerating the name would
be better of course but think about it twice because there's a huge cost to
doing that in the old model as I've explained above.
What I wanted to do here was to make progress since I had 2-3 days to work on
the new model again.
I see it's not easy and as a consequence we're not going to be able to
progress. It's just sad. I had the energy to code this quickly, modify
UserAvatarMacro to use the new model API and commit all that in master for
5.0M2… I don't have it anymore, nor do I have the time now (I had only 2 days
and they're now gone… :(…). It'll have to be someone else doing it.. or wait
till I get 10 free days in 1 year time to implement object naming in the
current model with all the breakages this would cause…
I had a good solution that was working both with the old model (class ref +
index - or numbers, whatever would work) and the new model (names when they're
implemented) and you're throwing it out… too bad...
Thanks
-Vincent
_______________________________________________
devs mailing list
[email protected]
http://lists.xwiki.org/mailman/listinfo/devs
