Hi, On 2019-02-02 11:29:10 -0500, Tom Lane wrote: > I think that the original idea here was that we should do as little > work as possible "up front", since most index paths will get discarded > before we reach createplan.c. But to the extent that that was valid > at all, it's gotten overtaken by circumstances. In particular, > postponing work to expand_indexqual_conditions (which is called by > create_index_path) is just stupid, because these days we typically > call that several times with the same index conditions. It's really > dubious that postponing commutation to createplan.c is a net win either,
It seems your approach isn't particularly in contradiction to the stated historical goal. We could create the new struct, but just not populate it eagerly, right? > Thoughts? If there's not objections I'd like to push this soon. Seems reasonable from a very very quick skim. Andres
