After doing some performance tests, I discovered that there is a huge I/O latency overhead introduced by Xen, so I've been working to move the OpenPisteMap database into the Dom0 (OpenPisteMap is run from a Xen DomU virtual machine). The move is now complete and database queries are definitely running a lot faster. The system is now (easily) keeping up with the minutely planet deltas. For now it is only running 2 render threads, but I will experiment with that setting a bit.

On a slightly different note, my initial attempt at working out which tiles are dirty after each update so that they can be automatically re-rendered turned out to be rather too simplistic. I had decided to try drawing a bounding box around every object and dirtying all the tiles within that bounding box, but unfortunately some objects can end up being huge (especially relations), so this resulted in an unfeasible number of tiles being marked as dirty. Some more intelligence is needed here - I think my next attempt will be to calculate which tiles a LINESTRING actually intersects with. I'm not entirely sure what to do with POLYGONs though - it may be wise to introduce some size limits so that polygons under that size limit have all the tiles within them dirtied whilst polygons over the limit would be treated as LINESTRINGs (the theory being that we probably aren't going to shade the area inside huge polygons).