On 20 January 2016 at 14:55, Konstantin Knizhnik <k.knizh...@postgrespro.ru> wrote:
> Hi, > > Hi, I glad to see that you interested in that too. >> I think this is a good feature and I think it will be very useful to have. >> I have already mentioned some related problems and possible improvements >> in my presentation. >> >> http://www.slideshare.net/AnastasiaLubennikova/indexes-dont-mean-slow-inserts >> Last two slides concern to this thread. Briefly, I've suggested to think >> about insertion buffer. Something very similar to it is already implemented >> in BRIN. It does not index last data from heap, while the number of last >> pages is less than pages_per_block. >> > > Do you mean GIN-like usage of insertion buffer (here it is called "pending > list")? > So that we have to combine search in the main tree and in the insert > buffer? > Actually this is what I want to avoided (because at least in case of GIN > pending list cause significant degrade of performance, > while up-to-date state of full text index is rarely required). Degrade in performance is because scan of pending list is O(N). If we did the same thing for monotonic inserts into a btree, the performance of ruling out any contents in the pending list would be O(1), so it is more feasible than you say. -- Simon Riggs http://www.2ndQuadrant.com/ <http://www.2ndquadrant.com/> PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
