JP <jeanpaul.ebejer@inhibox.com> writes:
> I have a 15-element/dimension vector (floats) data type. I also have about
> 10 million of these and given a query vector I would like to search these
> to find a number of nearest neighbours.
> For this I intend to extend the current implementation of kd tree
> in postgresql-9.2.1/src/backend/access/spgist in file spgkdtreeproc.c.
OK...
> I copied the spgkdtreeproc.c and created my own modified version (in
> the postgresql-9.2.1/src/backend/access/spgist) directory. How will
> postgres know when to use my functions instead of what is currently in
> place?
What you're missing is that you need to create an operator class that
links the query operators you want to support with these support
functions. For the built-in SPGIST opclasses, there are hard-wired
entries in src/include/catalog/ that set up the operator classes,
but for a user-defined extension you'd be wanting a script that does
CREATE OPERATOR CLASS. There are a number of examples for GIST and
GIN in contrib/ (but none as yet for SPGIST, IIRC). You could adapt
a GIST example pretty easily after reading
http://www.postgresql.org/docs/devel/static/xindex.html
> There are multiple sk_strategy defined (e.g. RTLeftStrategyNumber,
> RTRightStrategyNumber, RTSameStrategyNumber, etc). Who/where is this
> strategy being set?
Those are the numbers assigned to the operators by the operator class
definition. The C code has to agree with the CREATE OPERATOR CLASS
command about what number each operator has within the class, but
you can pick any numbers you want for an SPGIST opclass of your
own devising.
> I have read the chapters 52-54 of the postgres documentation but I am none
> the wiser - if you can point me to the right documentation/tutorial I would
> be eternally grateful.
Chapter 35 should help you a lot.
regards, tom lane