Re: Java OpenGroupware API - annotations

On Mittwoch, Sep 3, 2003, at 15:38 Europe/Berlin, Werner Schuster wrote:
> Hmm... one thing I noted  in this conversation,
> is the different conceptions about how high-level Enumerators 
> (Iterators)
> are;

Yes. Enumerators and Arrays are just concepts living at the same level 
of "abstractness" (Stack would be another one).
Yet in the actual case (records fetched using XML-RPC over high latency 
and slow links), enumerators match the actual data flow while arrays do 
not. So arrays should be build upon a stream (stream=enumerator).

> I personally think of Enumerator as more high-level than Lists; the 
> idea
> is that with an I don't have to think about what type of Container
> I am accessing; I simply call "next()" until nothing is left;

If you work on streamed data, the list needs to be build upon an 
enumerator. If you have indexed data, the enumerator needs to build 
upon a list. ;-)

Of course it depends on the underlying data structure which is a stream 
in case of HTTP. It may certainly be less obvious that XML-RPC results 
should also be treated as a stream - but for practical purposes it 
should be done that way.

(well, one could then claim that HTTP/XML-RPC then is the wrong base 
technology ;-) - this may even be true, yet SQL databases also work in 
a sequential way, so this matches pretty well)

> A List, for me, is lower level, as it is more similar to the actual
> memory layout (one element after the other) and I have to use a counter
> var to walk the list.

Of course. If you have a linked list, the enumerator will be more like 
the datastructure and indexed access be unnatural ;-) Now don't tell me 
that you never need linked lists ;-)

> So, in my eyes, the Enumerator (Iterator) is more of a convenience
> thing that most people will use;

No. Its just a different approach of accessing data.

> I have a problem with only offering an Enumerator and basing all
> accesses on it (eg. some convenience List class, that uses it);

You decide, I only want to bring up issues you'll run into ;-)

...image sample...
> But its a real pain in the neck, if you want to access a specific 
> pixel;

Yes, enumerators are usually more difficult to use. But scalable APIs 
*are* difficult to use ;-)
If network IO is as inexpensive as direct memory access and if we have 
100GB RAM machines, enumerators may not be worth the effort. But we 
have practical constraints.

This is getting theoretical, but interesting nevertheless ;-):

> I suppose, the programmers of this lib realized this and added some
> methods for indexed access (ie. something like getAt(x,y) );
> The problem was, they didn't access the data directly, but by way of
> the iterators... If you're not groaning with disbelief yet, think
> about this: if you want to access the pixel at pos. (100,100), you 
> have to
> have 100*100 iterations with the iterator...;  Now imagine the 
> performance
> of a program, that has to walk the pixel data from the bottom-right to
> the top-left pixel...

Whether or not this makes sense heavily on the application *and* on the 
data format in memory.
In case the whole image (eg consider a digital camera, high resolution 
image which easily has hundreds of megabytes) is really in memory as a 
1:1 matrix, direct access of course makes more sense. But if it is kept 
eg in RLE encoding to reduce the memory requirement (in RAM!) by factor 
50 - enumerators actually *do* make a lot sense.
Scanning the data with a GHz CPU will be way faster than consuming 
50times more memory.

> I know, that DB and Image Manipulations are different domains and whats
> bad for one need not be bad for the other, but thats one reason I am
> bit sceptic about using only Enumerators.

No, it pretty much shows the same issues. You are assuming that the 
image is kept completely in memory as a 1:1 matrix - but for a lot of 
digital media or print applications it often is not. And while 
providing indexed access may be more convenient, it is very likely to 
be much slower - often with results so slow that it can't be used in 
practice.

Its just about choosing the right algorithm/abstraction for the task at 
hand. If you only want to deal with 100 contact records in JOGI - 
indexed access is just fine and way easier (and probably faster as 
well). If you want to deal with larger databases, it still will be 
easier, but load times and memory requirements will be abysmal.

>> If you strictly focus on an OGo specific GUI client, this may be true.
>> In any other case it is not.
>> Mostly you will *not* keep them in memory - because you can't!
>
> Hmm... but thinking about OG.o ... would there really be such
> huge amounts of data? I mean, how many Person, Enterprise, ... objects
> are we talking about here?

I guess at least something like 100.000-1.000.000 objects. This 
obviously depends on the installation.

Note: the issue is not only about RAM (though thats probably the bigger 
problem). It is also about retrieval speed/API-latency over HTTP/XML.

>> JSP pages
>> - render object, forget object, render object, forget object
>> Export tools
>> - export CSV line, forget object, export CSV line, forget object
>
> True, it would be nice to work with a constant amount of data
> with one of these (basically just keeping the currently used
> elements in memory);

Of course.

>> Glow GUI Client
>> - map JOGI Person to SDBC Contact, forget object, etc
>> OGo GUI Client
>> - extract attributes you need (name,street,email), forget object
>
> Hmmm... I don't think so; for GUI applications, the  MVC approach would
> be used, so the Objects returned by JOGI would be used as models that
> store the data, and the GUI uses this data to display it; so, if you
> want to keep displaying the data, you need the stuff in memory;

That would be highly inefficient for larger datasets.

> Hmm... it seems, like there might actually be some need for ... hm...
> two APIs in JOGI; they would be like DOM and SAX for XML;
>  -the current JOGI implementation would be like DOM, which is easy to
> use, but memory intensive and unusable for big amounts of data;
>  - the other would only allow to process data serially as it comes in,
> but would use only constant memory;  but it would still offer 
> high-level
> objects (Account, Person,...) to easily access the data (so you don't
> have to fiddle with XmlRpc results);

Again, the comparison of SAX and DOM is excellent :-) DOM is just fine 
for small documents, but for big ones it doesn't scale.

But I do not agree that we need two separate APIs. The fetch API using 
the list can be easily layered upon the Enumerator one since this is 
natural (results coming in a streamed way, the list is being built).

> Although... offering a streaming API that really uses the advantages of
> streaming (ie. allowing access the first elements while the rest
> is still on the road) has its own set of problems;
> Well... actually only one problem: we would have to write our own 
> XmlRpc
> implementation, because currently available ones don't do that (at 
> least
> as far as I know); if you call a function that returns a list of 
> things,
> the method call returns when the *whole* list has been retrieved and 
> made
> available as a Java List;

Well, why break the JOGI API just because the initial backend 
implementation is broken ;-)


Anyway, I have made all my points. Decide now but don't complain later 
- after all its established practice that Java stuff is rewritten from 
scratch every six months ;-)

regards,
   Helge
-- 
OpenGroupware.org - http://www.opengroupware.org

-- 
OpenGroupware.org XML-RPC
[email protected]
http://mail.opengroupware.org/mailman/listinfo/xmlrpc