Re: AW: [sdo] ISSUE 99: Error in allocating SDO Type for elementswith n

Hi Frank, Blaise,

The idea of using default value in the definition of the behavior is very good. It maintains the difference between Int (default default value 0) and IntObject (default default value null) without bringing up instanceClass, finessing the language specific stuff nicely. And when the property definition gives another default value, then only xs:nillable properties can really be set to null, since otherwise we would change from null to default if there’s a round trip to XML.

Now I’m a little worried about the definition of nullable in section 4.5.5. This little box we are editing is the only thing that says how anything about setting property.isNullable, tho of course the attribute can also be set over the API. As Frank points out, section 4.5.5 currently says that setting non-nullable properties to null results in undefined behavior, in other words, we are saying it is a bad practice. Presumably, the proper way to accomplish something similar would be to call unset. The nice part of this rule is that the mapping between SDO and XML is clear: set to null means present in XML with xsi:nil=”true”, whereas unset means absent from the XML. At least in our implementation, if a property is not nillable then setting it to null is the same as unsetting it. Because for most java programmers, and certainly for anyone using static SDOs , calling unset is pretty unnatural. The spec specifically allows our behavior, so I don’t really have a problem. In 2119 terms, we have a MAY throw an exception or MAY do an unset. I’m wondering if we want to get rid of one of the MAYs, and replace the other with a MUST. That is, does anyone throw an exception in this case, or do we all do an unset?

I feel pretty strongly that we have to say something about the type of properties with xs:nillable=”true”. Metadata is one of SDO’s selling point, and saying that the only thing users can know about a nullable property is that it’s nullable seems hard. Perhaps a middle ground is to say that standard types (like Int) MUST be changed to the corresponding Object types (like IntObject), since this is the 2.1.1 behavior, and anyway, we don’t lose any information by doing it. But that User-Defined Types (like MySimpleType, below) do not change. This is clearly something that’s broken in 2.1.1, so I can’t believe there will be any compability issues.

Best Regards,

Ron

Von: Frank Budinsky [mailto:frankb@ca.ibm.com]
Gesendet: Freitag, 20. August 2010 16:44
An: Barack, Ron
Cc: sdo@lists.oasis-open.org
Betreff: Re: [sdo] ISSUE 99: Error in allocating SDO Type for elements with nillable=true

Hi Ron,

First, for the second sentence I think it would need to be:

A (Java) Implementation of SDO MUST return NULL as the property value for nillable properties having the xsi:nil=true attribute (and no other attributes), regardless of whether the instanceClass of the SDO type corresponding to the type used in the schema normally supports null values.

Even then, I wonder what this is really needed for. Why is it so important to dwell on the type/instanceClass? I think the spec is absolutely clear when a null value can be expected, and it has nothing to do with the type and whether or not the type is assignable from Object.class.

Setting the value of a property to null may or may not be allowed for a given property.
For example, a property that does not map to a nillable XML element or that maps to a
non-nullable RDB column, cannot be set to null. A property that can be set to null is
called a nullable property.
Calling get() on a property that is not nullable may still return a null value, if the default
value of the property is null and it is currently unset. Calling set(null) on a non-nullable
property will produce implementation dependent results. It may throw an exception or,
alternatively, it may cause the property to become unset.

From this very clear text, I'd say this is the test for expectedNull:

Boolean expectNull = pMstNillable.isNullable() || pMstNillable.getDefault() == null);

Notice that it has nothing to do with pMstNillable.getType(). In fact, the test that you provided:

Boolean expectNull = pMstNillable.getType().getInstanceClass().isAssignableFrom(Object.class);

is actually incorrect, according to the spec. It would incorrectly return true for a property that is not nullable and its default value is something other than null.

Your suggestion to change the spec to this;

Property name=[NAME]
type=[TYPE]
nullable=true

would break implementations that have two SDO types corresponding to the simple primitive types (e.g., XXX and XXXObject), and switch to the Object version in the nillable case. Personally, I think that's a bad legacy approach which was devised before we had the concept of Nullable, but changing it would be a breaking spec change.

Maybe the answer is simply to remove the (what I consider to be) confusing/ambiguous text but otherwise leave the table alone --- i.e., simply show that nillable maps to nullable:

Property name=[NAME]
nullable=true

Thanks,
Frank.

"Barack, Ron" ---08/20/2010 06:51:13 AM---Hi Frank, I think you understand what I was proposing.

From:	"Barack, Ron" <ron.barack@sap.com>
To:	Frank Budinsky/Toronto/IBM@IBMCA, "sdo@lists.oasis-open.org" <sdo@lists.oasis-open.org>
Date:	08/20/2010 06:51 AM
Subject:	[sdo] ISSUE 99: Error in allocating SDO Type for elements with nillable=true

Hi Frank,

I think you understand what I was proposing.

Let’s look at the text as it currently stands

If the type of the element has Simple Content without attributes, a Java Type with an Object instance class is assigned. For example, IntObject instead of Int.
In an XML document, xsi:nil="true" corresponds to a null value for this property.

I think we agree that the second sentence means:

A (Java) Implementation of SDO MUST return NULL as the property value for nillable properties having the xsi:nil=true attribute, regardless of whether the type used in the schema normally supports null values.

The disagreement is about the first sentence. I think you are correct is assuming “Java Type” should be “SDO Type”. I don’t believe that this mistake is so big that we can ignore the sentence entirely. And the example makes it perfectly clear that that the type of the property should have an instanceClass of IntObject and not int. Also, in the area above the text, where the translation of the XML is given, the Property’s definition does not contain a “type=[TYPE] like the other boxes do. The text is clearly trying to explain how the property’s type should be calculated. Given that, I would consider an implementation that says pMstNillable.getType()==MySimpleType to be broken *according to 2.1.1*. At least in my opinion. But let’s consider how things should be in 3.0

First of all, I think the behavior you want do define shouldn’t be defined in the chapter on the SDO-XSD mapping, but rather in the section (4.5.5) on Nullable properties…the user has no standard way of seeing if xs:nillable is true in the schema, but can easily check Property.isNullable(). The main use-case (that I can think of) where this would be relevant is where the user is inspecting the metadata, to determine if a null value is possible. Currently, this determination can be made based on Property.getType().getInstanceClass() alone, eg,

Boolean expectNull = pMstNillable.getType().getInstanceClass().isAssignableFrom(Object.class);

I think you want to change this so that the user needs to check not only the instanceClass, but also Propery.isNullable():

Boolean expectNull = pMstNillable.getType().getInstanceClass().isAssignableFrom(Object.class) ||
pMstNillable.isNullable();

I think this makes both Tuscany and your new implementation legal. I also think this is technically OK, even sensible, but I’m worried that it is this a breaking change. If every agrees that this should be the behavior, then I don’t think it will be hard to come up with appropriate wording in 4.5.5, and in the corresponding Java section. Then, the XSD-SDO mapping that we started with becomes simply

Property name=[NAME]
type=[TYPE]
nullable=true

without any text at all!

Have I correctly summarized the behavior you want to define? Like I said before, I need to think about whether this is a breaking change or not. Blaise, what do you think?

Have a nice weekend,
Ron
Von: Frank Budinsky [mailto:frankb@ca.ibm.com]
Gesendet: Freitag, 20. August 2010 00:03
An: sdo@lists.oasis-open.org
Betreff: RE: AW: AW: [sdo] AW: ISSUE 99: Error in allocating SDO Type for elements with nillable=true

Hi Ron,

Sorry, I did mean for my previous reply to go to the list.

In Tuscany we created an "Object" version of every user defined simple type which had a base class with a primitive instanceClass (just in case it's ever used with a nillable property). We would generate two SDO types from the XSD type: MySimpleType (instanceClass int) and MySimpleTypeObject (instanceClass Integer). Notice that's exactly the same pattern that was used with the built-in types (e.g., Int and IntObject). This is no coincidence, since a predecessor version of Tuscany was the primary implementation with which the SDO 1 specification was defined.

I think something like this is essentially what you're proposing (only not naming the Object version XXXObject, but rather making the Object types anonymous.

In another newer SDO implementation, we've handled this the way I've proposed below, simply by using the type itself, but allowing null to be returned as an "extended value", because the property is nillable. Because of the ambiguity in the spec, we concluded that this is really not noncompliant with the spec. IMO, the spec is only saying that the returned value's Java Type must be an Object type. Since get(primitiveTypeProperty) always returns the corresponding Object type anyway (e.g., Integer, since it can't return int) we're OK to also return null. Like I said in my earlier reply, the distinction between primitive and Object values really only comes into play in static SDO. This approach seems to be simpler than the Tuscany approach and works quite reasonably as far as I can tell. Can you explain an example of where this approach would really cause a problem?

That said, your point that I am saying the spec is currently ambiguous/broken but I'd like to leave it that way, is a good one. I'm certainly being a little irresponsible. However, if we do fix this in a way that will make the approach I've described noncompliant, I'd want to be convinced that there is really something wrong with our approach (i.e., it is really broken/bad) because if we change it, we would be in a position where our implementation will not be backward compatible if we want to be SDO 3 compliant. I rather not be in a position where backwards compatibility becomes an obstacle for moving to our implementation to SDO 3, if we can avoid it.

Thanks,
Frank.

"Barack, Ron" ---08/19/2010 05:09:26 PM---Hi Frank, You sent this to me, did you mean to send it to the SDO list?

From:	"Barack, Ron" <ron.barack@sap.com>
To:	Frank Budinsky/Toronto/IBM@IBMCA
Date:	08/19/2010 05:09 PM
Subject:	RE: AW: AW: [sdo] AW: ISSUE 99: Error in allocating SDO Type for elements with nillable=true

Hi Frank,

You sent this to me, did you mean to send it to the SDO list?

An old fashioned argument, just like the good old 2.1 days!

My opinion is that the spec is clear on the point that pMstNillable.getType().getInstanceClass() is (MUST BE) the instanceClass of IntObject. And, today at least, I think that’s pretty reasonable. Assuming I’m the one trying to get the spec to 2119 compliant, I’d probably have to put in a compliance point there. That doesn’t mean I won’t accept a resolution that loosens up the definition of compliant behavior, it just means that if we defer this, and you are unhappy with the defined behavior, we will just wind up discussing the same issue later. Let’s get it over with now. You even point out that you don’t know what the statement means, but then you go on to say you think the statement should be left as it is. Personally, I don’t even know if the statement belongs in Core or in Java. Let’s work out some wording here that everyone can be happy with.

I agree with Blaise that we are breaking SDO if we allow a property with instanceClass int, or deriving from Integer (as opposed to IntObject) to have null values. I think this use case is important, but not important enough to warrant such a fundament change to the relationship between data and metadata. Therefore, it seems pretty clear that pMstNillable.getType() cannot be MySimpleType. I think the spec should say so. We can maybe then get out of the issue about saying anything else about the type, leaving implementations free to use IntObject, an anonymous type, or anything else. I think the xmlType of the property should be the URI of MySimpleType.

In my reading of the spec, this is just adding the stuff about the XMLType.

Would you go with me so far?

Ron

From: Frank Budinsky [mailto:frankb@ca.ibm.com]
Sent: Donnerstag, 19. August 2010 16:14
To: Barack, Ron
Subject: Re: AW: AW: [sdo] AW: ISSUE 99: Error in allocating SDO Type for elements with nillable=true

Hi Ron,

The 2.1.1 spec does not say this:

"the property MUST be created with IntObject"

Mapping a property to IntObject instead of Int is simply an example of how to handle a nillable property of type xsd:int. All the spec really says about nillable elements is:

If the type of the element has Simple Content without attributes, a Java Type with an Object instance class is assigned.

I'm not even sure what a "Java Type" with an Object instance class means. Notice it doesn't say an "SDO Type" with an Object instance class. So that's the first ambiguous statement.

Let's assume that it really meant to say "SDO Type", then what is supposed to happen if the specified type, e.g., MySimpleType, doesn't have an Object instance class? The spec is very clear that unless sdoj:instanceClass is specified, the instanceClass of a type is the instanceClass of its base type, so I'd say you're example is an unsupported (error) case in the spec, since the type being requested cannot be used with a nillable property, because it's corresponding Java Type does not have an Object instance class.

On the other hand let's assume that all this means is that somehow, an implementation should figure out a way to arrange for the "Java Type" being used for the property, to be one with an Object instance class. With that reading of the ambiguous statement, all 3 of the our approaches (yours, Blaise's, and mine) would work.

Since it's been this way since SDO 1, I don't see any urgent need to try to fix it now. I'm happy to continue to leave the handling of this as implementation dependent.

Thanks,
Frank

"Barack, Ron" ---08/19/2010 09:11:15 AM---Hi Frank, In this case, deferring doesn't mean leaving things open, it means to keep the 2.1.1 behav

From:	"Barack, Ron" <ron.barack@sap.com>
To:	Frank Budinsky/Toronto/IBM@IBMCA
Cc:	"sdo@lists.oasis-open.org" <sdo@lists.oasis-open.org>
Date:	08/19/2010 09:11 AM
Subject:	AW: AW: [sdo] AW: ISSUE 99: Error in allocating SDO Type for elements with nillable=true

Hi Frank,

In this case, deferring doesn’t mean leaving things open, it means to keep the 2.1.1 behavior, i.e., the property MUST be created with IntObject, get(mstProperty) MUST return an Integer, never an int. And when you generate XSD from the MyComplexType, the <element> MUST have type=”xs:Integer”. Are you OK with this behavior, or can we do better? I agree we should change as little as possible, but I think we can at least say, starting with the 2.1.1 baseline, that the property should have the xmlType of MySimpleType, so that at least we do a better job of round tripping without breaking anything.

It seems a short step to create an anonymous type, which I think would really be a fix for the issue.

Ron

Von: Frank Budinsky [mailto:frankb@ca.ibm.com]
Gesendet: Donnerstag, 19. August 2010 14:38
An: Barack, Ron
Cc: sdo@lists.oasis-open.org
Betreff: Re: AW: [sdo] AW: ISSUE 99: Error in allocating SDO Type for elements with nillable=true

Hi Guys,

I don't think this issue is urgent and, from this note chain, we can conclude that this is not an issue for which we are close to consensus. I think we should defer it.

Thanks,
Frank.

"Barack, Ron" ---08/19/2010 06:50:09 AM---Hi Blaise, Frank, Three voices, three opinions is about what I expected. I think I'll go ahead and

From:	"Barack, Ron" <ron.barack@sap.com>
To:	Frank Budinsky/Toronto/IBM@IBMCA, "sdo@lists.oasis-open.org" <sdo@lists.oasis-open.org>
Date:	08/19/2010 06:50 AM
Subject:	AW: [sdo] AW: ISSUE 99: Error in allocating SDO Type for elements with nillable=true

Hi Blaise, Frank,

Three voices, three opinions is about what I expected. I think I’ll go ahead and add a fourth. First, some comments on your ideas:

@Blaise: I think the solution of making the instanceClass of all derived types Object types is certainly technically sound. My guess is that we are all storing values in Objects anyway. But I’m concerned about backwards compatibility. And I don’t think this is what users will expect.

@Frank: I think the difference in our perspectives boils down to the fact that I think it’s a problem for the instanceClass of a property to be different from the instanceClass of that property’s type. Property’s don’t have instanceClasses, Type’s do. It’s a pretty big change to suggest otherwise. Even if the static SDO issue falls away because we don’t generate the interfaces ourselves, there are certainly may other places where we assume that a property’s instanceClass is exactly what you’d expect from inspecting the type.

So, now on to another idea:

If we assume that for most applications, the correct representation of the DATA is more important than correct representation of the METADATA, then it is more important that get(pMstNillable) return null than it is that pMstNillable==pMst. If we want everything to be consistent , then that means that pMstNillable has to be or derive from IntObject. In order to round trip our metadata, we can add in that the xmlType of the property should be MySimpleType.

The spec currently says that pMstNillable.getType==MySimpleType. I think the only real problem with this is that we lose the facets or other metadata that is associated with the type. I think one possible solution to this would be to define an anonymous type that copies everything from MySimpleType other than its name and instanceClass. What’s good about this solution is that the nill values are represented AND the property’s values are consistent with the property’s metadata. And o it does not lose any metadata, as long as the user gets the type from property.getType and does not try to look it up over its name (which is already broken in 2.1.1).

Best Regards,
Ron

Von: Frank Budinsky [mailto:frankb@ca.ibm.com]
Gesendet: Mittwoch, 18. August 2010 22:54
An: sdo@lists.oasis-open.org
Betreff: Re: [sdo] AW: ISSUE 99: Error in allocating SDO Type for elements with nillable=true

I think MySimpleType having instanceClass int.class is fine. We just need to say that the instanceClass of a nillable property which is of a type that has a Java primitive instanceClass (e.g., int.class) will effectively be promoted to the Object version of the type's instanceClass (e.g., Integer.class) since null is also a valid value for the property. Effectively, this is just saying that the instnaceClass of a nillable property may not be exactly the same as the instanceClass of the property's type. Although the rule to figure it out is trivial, we could add a Property.getInstanceClass() method that returns it.

The value space of the pMstNillable property is the MySimpleType type's value space (i.e., MIIN_INTEGER to MAX_INTEGER) :

1. pMst unset
myObject.get(pMst) == 0
2. pMst set to 10
myObject.get(pMst) == 10
3. pMst set to nil
myObject.get(pMst) == 0 (or error since this is invalid data)

The value space of the pMstNillable property is the MySimpleType type's value space (i.e., MIIN_INTEGER to MAX_INTEGER) plus null (because it's a valid value for a nullable property).

1. pMstNillable unset
myObject.get(pMstNillable) == 0
2. pMstNillable set to 10
myObject.get(pMstNillabe) == 10
3. pMstNillable set to nil
myObject.get(pMstNillable) == null

Frank.

Blaise Doughan ---08/18/2010 04:11:08 PM---If the instance class for MySimpleType is *int.class* then I would expect: myObject.get(Integer.cl

From:	Blaise Doughan <blaise.doughan@oracle.com>
To:	"sdo@lists.oasis-open.org" <sdo@lists.oasis-open.org>
Date:	08/18/2010 04:11 PM
Subject:	Re: [sdo] AW: ISSUE 99: Error in allocating SDO Type for elements with nillable=true

If the instance class for MySimpleType is int.class then I would expect: myObject.get(Integer.class, pMstNillable) == 0
If the instance class for MySimpleType is java.lang.Integer.class then I would expect: myObject.get(Integer.class, pMstNillable) == null

The problem is we don't know what instance class to set on MySimpleType because we don't know it will be used by a nillable property. We could I guess in this situation always give these derived simple types the Object version of the instance class?

-Blaise

Frank Budinsky wrote:

Hi Ron,

Maybe I'm missing something, but the answers to me are straightforward.

Q1: What is the value of pMstNillable.getType()?
A1-1: pMst.getType(), that is pMstNillable.getType().getName().equals(“MySimpleType”);

Q2: Assuming “A1-1”, what should myObject.get(pMstNillable) return?
A2-2: null – this gives the detail about nil value, but I think users will still be surprised, since the value is inconsistent with the property’s metadata.
Why do you think this is inconsistent with the property's metadata? The property is "nillable", so a return value of null seems perfectly reasonable.

Q3: Assuming “A1-1”, what should myObject.get(Integer.class, pMstNillable) return?
A3-2: null – this is what users would expect. It would mean the implementation must remember the xsi:nill=”true” in addition to the integer value.
Remember what integer value? There is no integer value. The value is "nil". An instance can only have a value or nil, not both.

The only thing I think is special about nillable properties is that they need to return the Object version of primitive types in static SDO:

public int getMst() { ... }
public Integer getMst_nillable() { ... }

instead of the primitive instanceClass of the type, which won't allow null as a return value:

public int getMst_nillable() { ... }

Since the DataObject.get() methods always return Object types, there's no problem with them.

Thanks,
Frank.

"Barack, Ron" ---08/18/2010 09:33:49 AM---Hi Everyone, Based on yesterday's discussion, let me offer an example as a basis for discussion. Fo

From:	"Barack, Ron" <ron.barack@sap.com>
To:	"Barack, Ron" <ron.barack@sap.com>, "sdo@lists.oasis-open.org" <sdo@lists.oasis-open.org>
Date:	08/18/2010 09:33 AM
Subject:	[sdo] AW: ISSUE 99: Error in allocating SDO Type for elements with nillable=true

Hi Everyone,

Based on yesterday’s discussion, let me offer an example as a basis for discussion. For clarity, I’m leaving out all namespace stuff

We start with a schema that defines a simple type:

<xs:simpleType name=”MySimpleType” base=”xs:integer”>
<xs:minInclusive value=”5”/>
</xs:simpleType>

Let’s assume the schema defining this type is read in and that SDO has created the type. Sometime later, we read in another schema that uses the type in an element property… I’m imagining it in a second schema, if the use is in the same schema, we could potentially decide when defining the type that it’s instance class should be Integer and not “int”, but if the usage is in a second schema, then clearly we can no longer modify the existing type definition. So, let’s imagine a second schema

<xs:element name=”MyDoc” type=”MyComplexType”>
<xs:complexType name=”MyComplexType”>
<xs:sequence>
<xs:element name=”mst” type=”MySimpleType”/>
<xs:element name=”mst_nillable” xs:nillable=”true” type=”MySimpleType”/>
<xs:element name=”i” type=”xs:integer”/>
<xs:element name=”i_nillable” xs:nillable=”true” type=”xs:integer”/>
</xs:sequence>
</xs:complexType>

Here’s an instance document:

<MyDoc>
<mst>7</mst>
<mst_nillable xsi:nil=”true”/>
<i>21</i>
<i_nillable xsi:nil=”true”/>
</MyDoc>

Let’s imagine some code using these files:

DataObject myObject = xmlHelper.load(…).getRootObject();
Type complexType = myObject.getType();
Property pMst = complexType.getProperty(“mst”);
Property pMstNillable = complexType.getProperty(“mst_nillable”);
Property pI = complexType.getProperty(“i”);
Property pINillable = complexType.getProperty(“i_nillable”);

So, I guess the first question is:
Q1: What is the value of pMstNillable.getType()?
A1-1: pMst.getType(), that is pMstNillable.getType().getName().equals(“MySimpleType”);
I think this is what users would expect.
A1-2: sdo:IntObject, that is pINillable.getType().
This is what the spec currently says. But of course, we are losing information…in this case, the restriction. Not to mention that theres a round-tripping problem.

I guess A1-1 is preferable, but it creates problems when dealing with nils. This brings us to our second question:

Q2: Assuming “A1-1”, what should myObject.get(pMstNillable) return?
A2-1: 0 – this would mean ignoring the nillable attribute.
A2-2: null – this gives the detail about nil value, but I think users will still be surprised, since the value is inconsistent with the property’s metadata.
A2-3: A wrapper data object, and we could add some means of inspecting it for “isNull”… this is the pattern we use when dealing with dataTypes that need to be handled like objects.

Q2: Assuming “A1-1”, what should myObject.get(Integer.class, pMstNillable) return?
A2-1: Integer.valueOf(0) – this would mean ignoring the nillable attribute.
A2-2: null – this is what users would expect. It would mean the implementation must remember the xsi:nill=”true” in addition to the integer value.

BTW, sorry if there are errors in the schema or java code…this is only meant to be the basis for discussion, and I think the intent is clear.

My own current thinking is that the first answers to each question is the best choice. But maybe there are other solutions.

OK… let the discussion begin!

Ron

_____________________________________________
Von: Barack, Ron
Gesendet: Dienstag, 17. August 2010 16:31
An: sdo@lists.oasis-open.org
Betreff: ISSUE 99: Error in allocating SDO Type for elements with nillable=true

Hi everyone.

The problem description is here: http://www.osoa.org/jira/browse/SDO-99

In the current WD, the problem is in the table in section 7.4.2, in the block “ ;Element with nillable”. There are a lot of editorial issues with this text, for instance, the reference to Java types does not really belong in the core spec. but more substantively: the issue boils down to the problem that “xsi:nillable” is on properties, but “sdo:nullable” is on types. If you have a type that is defined as being an unboxed primitive value, we can’t simply change the type’s definition because it’s used in a nillable property. (I hope we all agree here).

In my opinion there are 4 possibilities.

1. We ignore the xsi:nillable
2. We use a SDO primitive object type (IntObject), and use the type referenced in the XSD as the XmlType only.
3. We use wrapper objects.
4. We defer the issue

I prefer option 3 and volunteer to write it up, if we get approval for the approach on today’s call.

Ron

sdo message