On 10/8/25 06:02, David Rowley wrote: > On Wed, 8 Oct 2025 at 08:15, Greg Sabino Mullane <[email protected]> wrote: >> I've been doing this sort of thing for clients a long time, and I always >> test both directions when I come across a query that should be faster. For >> real-world queries, 99% of them have no change or improve with a lowered >> rpc, and 99% get worse via a raised rpc. So color me unconvinced. > > I wonder how much past experience for this on versions before v18 > count in now that we have AIO. The bar should have moved quite > significantly with v18 in terms of how often Seq Scans spend waiting > for IO vs Index Scans. So maybe Tomas's results shouldn't be too > surprising. Maybe the graph would look quite different with io_method > = 'sync'.. ? >
Interesting idea, and I'll try to run this on 17 and/or on 18/sync. I should have some results tomorrow. But based on the testing I've done on 18beta1 (in the thread about what should be the default for io_method), I doubt it'll change the outcome very much. It showed no change for indexscans, and seqscans got about 2x as fast. So the random_page_cost will be about 1/2 of what the earlier results said - that's a change, but it's still more than 2x of the current value. Let's see if the results agree with my guess ... regards -- Tomas Vondra
