Excerpts from Nina Jeliazkova's message of Mon Nov 16 12:10:52 +0100 2009:
> Hi Tobias,
> 
> > Hi everyone,
> >
> > On Fri, 2009-11-13 at 13:49 +0200, Nina Jeliazkova wrote: 
> >>
> >>
> >> 2)  Does the proposal means we abandon the API that allows to retrieve
> >> feature values, given a compound and feature identifiers ?
> >
> > Another question: Does the proposal imply that features are coupled now
> > to datasets? That would mean, that we cannot have a compound with
> > features stored, that is not in a dataset? Or am I missing something? 
> >
> 
> This is what was left without discussion (IMHO). I am not sure this is a good
> option, there are lot of compound properties which are independent of any
> dataset.
> 
> >
> > If I calculate a descriptors with the new API...do I update the dataset
> > or do I create a new one? The latter might lead to a huge number of
> > datasets and maybe even redundancy. 
> 
> Exactly.  In addition, it will make more difficult combining features for the
> same compounds from different datasets, which is crucial for some use cases as
> read across.
> 
> 
> From my point of view compounds are separare entities, compounds have features
> and datasets are purely for denoting subsets of compounds and features.  Thus
> my disagreement with the proposal to abandon feature API.
> 

Maybe I got it wrong, but my understanding from our Munich discussion
was, that compounds and their features are generally accessed through a
dataset service.

It seems that your intention was to access the features of compounds
through the compounds service and compounds of features the feature
service and use the dataset service merely for subsetting data. Is that
correct?

I would like to have a single service for accessing compounds and
features. This does not necessarily mean, that every
compound-feature-value triple has to belong to a dataset_id.  We can
provide e.g. the same methods for /dataset as for /dataset/{id} or move
the dataset API to the compound service and make dataset_uri a query
parameter. 

I would however opt against replicating more or less similar methods in
the compound, feature and dataset APIs.

Best regards,
Christoph