Jerry Sievers <jerry@jerrysievers.com> writes:
> Tom Lane <tgl@sss.pgh.pa.us> writes:
>> What I still don't understand though is why you see it
>> in 8.1 and not 8.0 ... the src/port/qsort.c code didn't change at all
>> between those versions. Maybe 8.0 isn't taking the qsort code path,
>> perhaps because it uses a shade more memory or some such. Could you
>> try increasing maintenance_work_mem even more, like to 100M,
>> and see if 8.0 gets slow?
> I did as you suggest here; taking the MWM setting incrementally up to
> nearly a Gig in 100M increments, also raised and lowered the work_mem
> setting too.
> Tried a dozen or more combos and in every case, the index built
> quickly on 8.0.3.
On further analysis, it seems the problem is dependent on the exact
ordering of the inputs to the qsort function. So not only do you need
maintenance_work_mem to be large enough that the code will try to use
qsort, but the physical order of the rows in the table matters.
I suspect that you are testing on an 8.0 table with a different physical
row order --- if you drop the table and reload it from the same dump you
loaded into 8.1, does it get slow?
regards, tom lane