[DAS] Adjacent feature extension

[Thomas Down]

On Mon, Mar 7, 2011 at 10:04 AM, Andy Jenkinson <andy.jenkinson at ebi.ac.uk>wrote:

> Hi Thomas,
>
> Thanks for this. Regarding the option of whether to return just one feature
> per side or all overlapping features, the only other advantage that
> immediately springs to mind for the latter (in addition to some measure of
> consistency, as you mention) is that it allows the client to immediately
> render the exact region of that feature without triggering another request.
> It would generally mean changing zoom level. I'm can't say if clients are
> likely to follow this mechanism as opposed to, say, pan and centre on the
> feature, but if they wanted to it would be more efficient (and possibly a
> little bit more efficient anyway depending on how your client does its
> requests).
>

Yep, I agree.  I'd be interested to learn whether there are any clients that
would seriously consider taking advantage of this.  My own thinking is that
even if we do adjust zoom level (as Dalliance sometimes does, e.g. in the
"jump to gene..." navigation op), clients are much more likely to zoom to a
view that contains the target feature plus a "sensible" amount of flanking
sequence, rather than a view where the target feature is perfectly framed.

Furthermore, this rather seems like optimizing for the case where only one
annotation source is active.   Surely we're talking about the
*distributed*annotation system, and clients will still have to go off
and query all the
other annotation sources, even if they are able to skip the one which
responded to the "adjacent" query.  So long as there's some kind of query
parallelization in place, this probably isn't a performance issue.

Do any other client developers feel differently?


> Disadvantages I can think of:
>  - "adjacent" request takes marginally longer
>  - not quite as obvious what clients should put in their UI controls - need
> to pick a feature to be able to do "jump to BRCA1"
>  - risk of servers not implementing it correctly and only returning one
> feature anyway (although I don't think this is likely as the concept is
> different to "feature-by-id")
>
> Some things to further define:
>  - servers can't return a fake feature
>

Yep, will clarify this.


>  - should servers return features on different reference sequences if there
> are none one the current one?
>

In my opinion, absolutely yes.  Otherwise the "10 features in the genome"
case remains a massive pain (and potentially a disaster, for
inhomogeneous-dstributed data; won't someone think of the MHC tiling arrays?
:-).  And even worse for the "10 features in UniProt" case (where I can also
see this feature being quite interesting).

I've tried to be explicit about this in my proposal (see the penultimate
paragraph + example 3), but any suggestions for further clarifications are
welcome.


>  - how should servers treat features that overlap the adjacent range? Treat
> them as the adjacent feature to return, or only include features completely
> outside the query range? What if the next feature completely outside the
> query range is part of the same feature hierarchy (e.g. an exon outside the
> current window).
>

It's a point rather than a range, but yes I agree this is still an open
question.  I'd actually written the spec such that overlapping features do
get returned (on the assumption that clients will do "trivial" cases of
next/previous feature in-memory without a network round trip), but again if
other client developers do things differently, I'd like to know.

I think "include overlapping" will have less special-cases to worry about,
though.  e.g. the PART/PARENT issue you allude to.  Let clients deal with
that ("dumb servers, smart clients").

                 Thomas.