Clustering Spatial Data to minimize disk I/O

Question

Hi,
I use Oracle Spatial on an Oracle RDBMS 11gR2 EE.
I have read that wiht huge spatial tables (few GB for instance), it is possible to "cluster" the rows having an SDO_GEOMETRY data type.
This can minimized I/O, having spatial geometries grouped in contiguous data blocks.
I have also found that there is the function MD.HHENCODE encodes a two-dimensional point ino a raw string and this string can be used to sort/re-organize the points in a table.
I have not find how this function works. Does someone know the principle on which this function works?
How does this apply to polygons or lines, except by conputing a centroid for these geometries?
I don't know if I get good approach, but generally speaking, does someone know how to "cluster" spatial data to minimize I/O?
Thanks by advance for any tip.
Kind Regards.

Answer 1

Hi Unnamed Cat Person,
Sorry I am late responding to your questions.  You should sign your postings with the name of your cat.  I pester posters to provide a name not just to be petulant, but unless we have names to reference it can be difficult to tell if I am responding to you or to Stefan or to Bryan.  It certainly does not have to be your real name, the cat's name would work just fine.
So you are looking for information not on how to do spatial sorting but rather the "why to do this" and "how does it work" questions surrounding spatial sorting.  This forum is generally not a deep well of such information as it tends to focus on practical implementations.  As a group we often should be more introspective on the reasons why things work they way they do and if there are better ways to do them but then to some degree that is what our employers pay Oracle to do.  You might get better answers for such questions over on stack exchange (Geographic Information Systems Stack Exchange) or you might not.
MD.HHENCODE is an undocumented, fully wrapped, by default unexecutable function in Oracle Spatial - so you won't find anything on this from Oracle.  Exactly how it works is "beyond our ken".  The 12c version of Pro Oracle Spatial is due out next year and I personally wonder if they should keep this section in the book.  I should also note its quite the pain in the backside to use returning a RAW datatype of all things.  So as the book says, the function "more or less" encodes a point in spatially sorted order.  So its very difficult for any of us to say if this function works well or sensibly or follows some established formula, etc.  Its a proprietary black box.
So taking a step back and cribbing from Simon's discussion, you also could just sort on the X and Y values.  This is what PostGIS "order by geometry" does and you could set up something similar in Oracle quite easily.  But I don't really see much value in this approach if you are trying to cluster things together.
Now Simon's article says that sorting on the Morton key is better.  Does he say why Morton keys are better than the above methods?  I guess he does not have a summary section at the top but I feel the results speak for themselves, particularly the illustrations.  Imagine this article without those illustrations and it would be most difficult to visualize.  There are also good illustrations on the wikipedia page
Z-order curve - Wikipedia, the free encyclopedia
Note that as none of us can say how HHENCODE functions, none of us can really state definitively that Simon's Morton Curves are better.  Oracle could quietly change the function tomorrow to produce Morton curves (for all I know it does already as I haven't used the thing in years).  But I think Simon's work is fantastic so in my opinion then his way is better!  Take that for what its worth.  Show his nifty illustrations to your boss and that will end any discussion of your rational.
But wikipedia suggests that an even better way to go is using Hilbert curves.  So hey, does anyone on the list have this implemented in oracle spatial?
Hilbert curve - Wikipedia, the free encyclopedia
So there you go, four approaches:
X Y striping = useless for clustering
HHENCODE = "more or less" works okay in unknown manner
Greener Morton Curves = works well, fully explained, coded for you
Hilbert Curves = might be better, fully explained, you need to code it yourself
Does anyone on the forum have more entries to add?
If you are heck-bent on spatially sorting your data, I'd suggest using Simon's code.  You could also do a head-to-head test against HHENCODE and report back to us your findings.  My suspicion though is that as Bryan and Stefan both allude, that spatial sorting won't cure the problem that brought to you to the forum.  I am not saying there could not be a modest little boost, but its dependent on many other factors I mentioned in my first post.  Anyhow, I have appreciated your posting, it lead me to think more about a topic I've not paid much attention to for years.  We could spend more time in this forum on such issues.  But most often things here are just about how to do a task or bug chasing.  Perhaps there is a better place for these things.  Maybe if a question leads beyond the "how" and "bug" topics, we should cross post into stack exchange?  I dunno.
Cheers,
Paul