"Select * " on 12-18M row table from remote machine thru JDBC - Performance nose-dives after 10M-ish records

On machine 1 - a table that contains between 12 and 18 million rows
On machine 2 - a Java app that calls Select * on the table, and writes it
into a Lucene index

Originally had a fetchSize of 10,000 and would take around 38 minutes for 12
million, 50 minutes for 16ish million to read it all & write it all back out
as the lucene index

One day it started taking 4 hours. If something changed, we dont know what
it was

We tracked it down to, after 10 million or so rows, the Fetch to get the
next 10,000 rows from the DB goes from like 1 second to 30 seconds, and
stays there

After spending a week of two devs &  DBA trying to solve this, we eventually
"solved" it by upping the FetchRowSize in the JDBC call to 50,000

It was performing well enough again for a few weeks

then...one day... it started taking 4 hours again

we tried upping the shared_buffer from 16GB to 20GB

And last night... it took 7 hours

does anyone have ANY ideas?!

we are on postgres 9.1

the Server merely shows an idle transaction, waiting to be asked for the
next bunch of rows - not IO or CPU bound

thanks much




--
View this message in context:
http://postgresql.1045698.n5.nabble.com/Select-on-12-18M-row-table-from-remote-machine-thru-JDBC-Performance-nose-dives-after-10M-ish-records-tp5725854.html
Sent from the PostgreSQL - jdbc mailing list archive at Nabble.com.