David Gould <da...@sonic.net> writes: > I also thought about the theory and am confident that there really is no way > to trick it. Basically if there are enough pages that are different to affect > the overall density, say 10% empty or so, there is no way a random sample > larger than a few hundred probes can miss them no matter how big the table is. > If there are few enough pages to "hide" from the sample, then they are so few > they don't matter anyway.
> After all this my vote is for back patching too. I don't see any case where > the patched analyze is or could be worse than what we are doing. I'm happy to > provide my test cases if anyone is interested. Yeah, you have a point. I'm still worried about unexpected side-effects, but it seems like overall this is very unlikely to hurt anyone. I'll back-patch (minus the removal of the unneeded vac_estimate_reltuples argument). regards, tom lane