Poul Nielsen wrote:
> I think that the best policy is to evolve CellML toward a clean and  
> simple specification. I don't think that this means that we require a  
> complete break with previous specifications at each major iteration  
> if, for example, we use deprecated/obsolescent flags. I believe that  
> it is essential, however, that we provide a mechanism for adding and,  
> importantly, removing elements and attributes.
>   

I agree that there are circumstances when elements and attributes can be 
added or removed. However, the impact of this depends greatly on which 
elements are being added or removed.

> We have discussed previously the option of retaining deprecated/ 
> obsolescent elements/attributes for one or several iterations with a  
> view to maintain compatibility, but signaling that more appropriate  
> constructs are available and that such features are marked for  
> deletion. Examples of what is at issue here are suggestions that (1)  
> the reaction element be removed 

This fits into the category of changes which would probably not have a 
major impact on the current software base, because (as far as I know) 
there are no tools that implement reaction correctly at present. This is 
contrasted with a tangible benefit to removing reactions, specifically, 
that compliant software does not need to implement it (and historically 
this has been a cause of deviation from the specification).

> and (2) the public_interface and  
> private_interface elements be removed or their attributes be  
> modified.

This latter case, on the other hand, fits into the category of changes 
which would break every single existing CellML 1.1 model and force it to 
be regenerated for CellML 1.2. I see this as having a much greater 
impact than removing reactions. On the other hand, while the 
directionality on interfaces is unnecessary, it is not a huge 
implementation burden - we could just maintain directionality as a 
historical quirk of CellML.

I think it is a question of taking a pragmatic vs an idealistic 
strategy. The idealistic strategy is to break compatibility on even the 
most simple models to clean up the specification, while the pragmatic 
approach will try to balance compatibility against specification 
improvement. I would argue that for the best strategy for the long term 
future of CellML is to be somewhat pragmatic, and try to avoid forcing 
every single model to  recoded.

I think that this approach is also very common as programming languages 
evolve. For example, programs can be written which are valid in 
FORTRAN90 and FORTRAN77. Most valid FORTRAN77 programs would be valid 
FORTRAN90 programs, although there are some F77 programs which are not 
valid F90 programs. If you wrote a program in F90 and didn't use any F90 
features, it should be a valid F77 program. Likewise, a similar 
situation can be seen with the various revisions of C++ (and even 
between C and C++).

>  Deprecation has the advantage of offering a more gentle  
> transition, allowing models expressed in older, but not every,  
> iteration specification to be interpreted. It has the disadvantage  
> that it may not provide a clean break, making it difficult to deal  
> with the addition of new features that may be incompatible with old  
> ones. In both scenarios it will be important to provide tools to  
> translate older models into newer versions that conform to later  
> specifications.
>
> Because I think that it is important to be able to remove elements/ 
> attributes, I am not in favour of option A. Like Mike, I have a  
> problem with the statement that "[future versions] may remove  
> functionality that does not have an established base of software  
> which correctly implements that functionality". If the addition or  
> removal of elements or attributes results in a better specification,  
> then I think that consideration should be made for such changes  
> irrespective of whether there is "an established base of software  
> which correctly implements that functionality".
Although the same argument could be used, for example, to suggest that 
C++ should clean up the exceptions or any of several other well 
documented syntactic issues which C++ has inherited from early versions 
or from C, and yet the committees for these languages tend to avoid 
changes which break compatibility too much. While CellML isn't exactly a 
programming language, I think we should probably take note of the way 
they handle compatibility.

>  My preferred option  
> is C, with the proviso that translation tools and APIs, conforming to  
> the new specification, be made available at the time of transition.
>   
I think that realistically, tools will take time to develop, and 
although we could go through an extended draft phase as we have done in 
the past, people will use the new format as soon as tools support it, 
and we should probably try to minimise the level of incompatibility this 
causes (subject to our need to add new features), even if we are stuck 
with the ugliness of some historical decisions for longer because of it.

Best regards,
Andrew

> Best wishes
> Poul
>
> On 2008 Jan 09, at 11:58, Andrew Miller wrote:
>
>   
>> Hi all,
>>
>> There have recently been some discussions of changes which would
>> drastically break forwards or backwards compatibility of CellML (for
>> example, changing the way that connections work).
>>
>> I think that it is important that we come to some consensus on what  
>> the
>> policy for inter-version compatibility in CellML should be quite soon,
>> because this drastically affects decisions that need to be made in
>> CellML specification development.
>>
>> It doesn't really make sense to be inconsistent with respect to  
>> version
>> compatibility - it would be quite unfortunate if we worked hard to  
>> keep
>> compatibility for one part of CellML, and then broke it in another  
>> major
>> part such as by changing the way connections work, and so I think we
>> need a policy on this.
>>
>> I have come up with a number of different potential policy  
>> statements on
>> when backwards compatibility should be broken and when it should be
>> kept. It might help us to reach consensus if members of the CellML
>> community could rank the policies in order of preference (1 is the  
>> most
>> preferable policy, 2 the next most, and so on), and suggest any good
>> policies that may be missing.
>>
>> Option A)
>> Future versions of CellML should aim to solely express the  
>> intention of
>> previous versions better and more clearly. They should aim to keep  
>> full
>> compatibility with an implementation of the specification according to
>> the rules of the specification as they were interpreted by  
>> implementors.
>>
>> Option B)
>> Future versions of CellML should try to be mostly compatible with
>> existing implementations of previous versions of CellML. They may  
>> remove
>> functionality that does not have an established base of software which
>> correctly implements that functionality. They may also add in new
>> functionality, if that new functionality significantly increases the
>> expressiveness of the language. However, in normal circumstances,
>> compatibility should be maintained, so that when a model not using new
>> features is saved in the new version's most preferred format, it can
>> still be correctly loaded into software only supporting the old  
>> version.
>> Likewise, a model not using any removed features should be able to be
>> loaded in software supporting only the new version of the  
>> specification.
>>
>> Option C)
>> Future versions of CellML should make any changes which make it
>> conceptually cleaner, even if there is a less clean compromise  
>> available
>> that would have lesser compatibility implications. Software will  
>> need to
>> explicitly support more than one version as a completely separate  
>> format.
>>
>> My preferred choice is Option B. Despite being apparently at opposite
>> ends of the spectrum, Option A and Option C are, in my opinion, fairly
>> similar, because if we adopted Option A, larger changes would  
>> appear in
>> a new specification called something other than CellML. Although there
>> could be advantages of coming up with a more meaningful name than
>> CellML, I think that this would also set us back in terms of community
>> awareness of the specification, and so I think that Option C is
>> marginally better than Option A (i.e. my personal order of  
>> preference is
>> currently B:1, C:2, A:3).
>>
>> I look forward to any feedback on this you may have.
>>
>> Best regards,
>> Andrew
>>
>> _______________________________________________
>> cellml-discussion mailing list
>> cellml-discussion@cellml.org
>> http://www.cellml.org/mailman/listinfo/cellml-discussion
>>     
>
> _______________________________________________
> cellml-discussion mailing list
> cellml-discussion@cellml.org
> http://www.cellml.org/mailman/listinfo/cellml-discussion
>   

_______________________________________________
cellml-discussion mailing list
cellml-discussion@cellml.org
http://www.cellml.org/mailman/listinfo/cellml-discussion

Reply via email to