On Fri, May 1, 2015 at 5:27 PM, Andrew Gierth
<andrew@tao11.riddles.org.uk> wrote:
> Also as I've pointed out, it's not even clear that there is a regression
> at all, since I've already shown that changes of several percent in
> timings of sort operations can be caused by irrelevant noise factors.
> To actually show a performance regression of less than 10% or so would
> require, at a minimum, showing two different timings using the same data
> and the same binary, though even that is subject to noise; to really
> prove it you'd have to show a statistically significant difference
> between sets of binaries with random padding sizes (see the graph I
> posted on this point).
I think the issue is somewhat confused by the fact that there was
performance investigation work done on the thread, and a regression
was investigated (a regression that has since been fixed). This was a
problem that had nothing in particular to do with the Datum tuplesort
abbreviation patch, though.
--
Peter Geoghegan