Spatial index support via Pg12 support request API #378
Conversation
|
Wow, i was tracking this but i missed it getting merged upstream. Do you mind keeping this open for a few days / a week so I can have a look and setup some tests, please?
With this I'm all for updating the cost of all functions and telling people they should use PG12 for proper cost. OTOH, lately I've been thinking about how could Postgres support costs based on input (for example St_Area costs 5*avg points per geometry in input table squared). Do you know if there is something like that in the horizon (either for PG12 or 13) where we can help/push? |
|
Yes, dynamic costing is available now However, in order to be truly dynamic I think we'd need to enhance our stats tables to not only store grids of feature density but also grids of vertex density. Or even just save a single slot for the average vertices/feature. Or I guess maybe there's already enough data available in the stats slot... (object size and number of objects)... can you tell I haven't looked at the stats system in a while? |
supportfn to geography inlines
raster/topology tests... no dice, there's still regression failures in topology
leaf nodes... fixes topology issue sqlmm, but there are *still* topology regressions!
|
@strk I really need your help figuring out exactly what st_function(geometry,geometry) is causing the topology failures... something that breaks it down to an actual "these inputs are returning an incorrect answer" situation. I cannot tease that out of topology. https://travis-ci.org/postgis/postgis/jobs/500590893#L2870 |
now the topology tests pass.... there's something worrisome about this, but I'm not sure what.
rather than ignored (or, so that the real problem can be exposed; either way)
Using the support request API it is possible to add index support to functions without recourse to SQL inlining. So, we can add an "overlaps" index condition to the plan for a query that has
ST_Intersectsin theWHEREorJOINclause, invisibly, by implementing an appropriate "support function".A side effect of moving to this means of activating indexes is allowing us to re-cost functions with higher cost, but without causing index failures due to inlining failing, as we have seen in the past.