[postgis-devel] Re: [HACKERS] GiST memory usage

[Mark Cave-Ayland]

> -----Original Message-----
> From: strk at refractions.net [mailto:strk at refractions.net] 
> Sent: 16 August 2005 12:29
> To: Martin Daly
> Cc: Paul Ramsey; Mark Cave-Ayland (External); 
> postgis-devel at postgis.refractions.net
> Subject: Re: [postgis-devel] Re: [HACKERS] GiST memory usage
> 
> 
> I've done a pass on memory usage in GiST indexing.
> It doesn't seem we're really leaking, anyway, in the unlikely 
> event your tables are full of EMPTY or INFINITE geometries 
> the change should reduce memory usage.
> 
> I'll be interested in hearing from you Martin about box-only 
> version of the index. Does it still hit memory limits when 
> DETOASTED values are smaller ?
> 
> Teodor Sigaev (GiST developer) reports he's been able
> to build rtree indexing on 1e9 records...
> 
> --strk;


Hi strk,

I had a quick look at this yesterday and think I found a small leak in
LWGEOM_gist_picksplit. It's around line 1248 when the split direction is
ambiguous - interLR and interBT are generated by BOX2D_intersects which
pallocs() a new BOX2DFLOAT, however they are not pfreed() after the sizeLR
and sizeBT comparison.

I've loaded in a couple of small shapefiles, but in those cases the
ambiguous split code doesn't seem to get called even once. So unless with
very large datasets the ambiguous split case occurs more often, I can't say
that this will solve the problem entirely. However, I could provide a
replacement liblwgeom.dll for Martin to test both our fixes if we can't find
anything else.


Kind regards,

Mark.

------------------------
WebBased Ltd
17 Research Way
Tamar Science Park
Plymouth
PL6 8BT 

T: +44 (0)1752 797131
F: +44 (0)1752 791023
W: http://www.webbased.co.uk