[Qgis-developer] Potentially serious performance regression in new geometry - should 2.10 be delayed?

[Nyall Dawson]

Hi all,

Unfortunately, we've become aware of a serious performance regression
caused by the new geometry engine. Basically, the situation is that for all
geometry operations which rely on geos (think buffers, splits, spatial
relation operations such as intersects and within,... ) the geometry now
needs to be converted into a geos representation with *every* operation. In
the old geometry engine this conversion was done once, and the result
stored so that follow up operations would not need to recalculate it. This
potentially has huge impacts on the performance of common tasks such as
selecting all features which intersect a geometry.

I've had a look, and unfortunately it's not trivial to fix this. I think
the correct solution to this is to:

- make QgsGeometryEngine accept and return QgsGeometry containers, not
QgsAbstractGeometryV2
- store the generated geos representation of geometries within
QgsGeometryPrivate inside the QgsGeometry container. This way it will be
reusable between different geometry operations, and shared when QgsGeometry
objects are copied. This will also have the benefit that if a geometry is
prepared using geos then subsequent geos operations performed on that
QgsGeometry and its shared copies will be much faster.
- make QgsGeometry a friend class of QgsGeo, so that it can access
QgsGeometryPrivate to retrieve or set the geos representation of the
geometry as required

An alternative (short term) solution would be to just cache the geos
representation when geometry operations are called through the older
QgsGeometry modification/relationship operations. This would be easier, but
means that the API of QgsGeometryEngine will be stuck with the current
design, and we won't be able to properly fix this until breaking the api
for 3.0.

Either way, I doubt this can be addressed within the remaining 3 days we
have until release. Should we delay to address this? Release with the
regression? Or am I missing something and there's an easier solution we
could implement? Or even possibly this additional cost of recalculating the
geos representation is trivial and can be ignored (maybe someone could test
this with a little repeated intersection script)?

Thoughts?

Nyall
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osgeo.org/pipermail/qgis-developer/attachments/20150623/009f2cdb/attachment.html>