Le 2013-04-17 à 14:15, Jeff Janes a écrit : > On Wed, Apr 17, 2013 at 4:26 AM, François Beausoleil <franc...@teksol.info> > wrote: > > > Insert on public.persona_followers (cost=139261.12..20483497.65 > rows=6256498 width=16) (actual time=4729255.535..4729255.535 rows=0 loops=1) > Buffers: shared hit=33135295 read=4776921 > -> Subquery Scan on t1 (cost=139261.12..20483497.65 rows=6256498 > width=16) (actual time=562265.156..578844.999 rows=6819520 loops=1) > > > It looks like 12% of the time is being spent figuring out what rows to > insert, and 88% actually doing the insertions. > > So I think that index maintenance is killing you. You could try adding a > sort to your select so that rows are inserted in index order, or inserting in > batches in which the batches are partitioned by service_id (which is almost > the same thing as sorting, since service_id is the lead column)
In that case, partitioning the original table by service_id % N would help, since the index would be much smaller, right? N would have to be reasonable - 10, 100, 256, or something similar. Thanks, François
smime.p7s
Description: S/MIME cryptographic signature
