2008/10/14 Robert Treat <xzilla@users.sourceforge.net>
>
> On Monday 13 October 2008 04:53:44 Markus Wanner wrote:
>
> > Having reviewed the last commit fest's intagg patch as well, I thought
> > we agreed that a more general functionality is wanted for core. But as
> > long as we don't have that, I'd like intagg to stay in contrib.
> >
>
> While I agree that the "right" solution would be to make this code work more
> generally for other aggregates, I also think that until someone is willing to
> do that work, this needs to stay in contrib, and that we ought to accept
> patches improving it.
I started to look at implementing array_agg by making the existing
intagg stuff more generic, but I came across an issue (which might
also be a bug in intagg). Intagg relies on being able to stomp on the
input transition value, and basically allocates the working array as a
block, reallocating it when it becomes too small. The lower bound of
the array is (ab)used to keep track of how many items have been
allocated.
For a generic aggregate, which accepts any types and NULL inputs, in
order to avoid a linear search through the data to find where to put
the next element, it would be useful to instead store the offset to
the next free byte in the lower pointer.
The problem is that I can't see any way to guarantee that someone
wouldn't create another aggregate using the same transition function,
but giving an initial value to the lower bound which will cause the
transition function to do naughty things (as far as I can tell, this
is also true of intagg - giving an inital state value of
'[200000:200000]{1}' will cause it to happily write up to 200000
integers off the end of that one element array without allocating any
extra storage...
I'm not sure what the best way around this is - it seems that
performancewise, avoiding an array_seek() call in the transition
function would be good. Is there some way that the transition function
can tell which context the state value was allocated in, and thus
whether it was supplied as an initial value or not?
Ian
