FW: BTP_CHAIN flag was expected
От | AliE@atlas.com |
---|---|
Тема | FW: BTP_CHAIN flag was expected |
Дата | |
Msg-id | E5B27B6CBA34D111B8FC0000C0105CF96ABC9B@ntserver3.atlas.com обсуждение исходный текст |
Ответы |
Re: [HACKERS] FW: BTP_CHAIN flag was expected
(The Hermit Hacker <scrappy@hub.org>)
|
Список | pgsql-hackers |
Here is the latest I received from our Field Engineering regard to the Index Corruption and the growth of our tables. I would appreciate any help to figure out why we get the index corruption and why our tables grow so fast? [Ali Ebrahimi] alie@atlas.com > > PG_VERSION v6.3, FreeBSD 3.0-971031-SNAP > > The 'before and after' file lists below show a fifty percent > increase in the size of user account during the course of one > day. > Since the database is around 300,000 records we might assume that > the > increase reflects about 150,000 added records. We are averaging > about > 25,000 completed transactions to acct_history each day. This > data > suggests about six updates to user_acct for each update to > acct_history which is about what we expect. > > What we don't expect is for our user_acct to grow (so much!) in > size > with simple updates. > > We also don't expect 'btree: BTP_CHAIN flag was expected' > errors to > be popping up. We have seen btree errors in both acct_history > and > user_acct. acct_history_acct_no_idx is non-unique, > user_acct_card_no_idx is unique, the other user_acct indexes are > non-unique. All indexes are btree. > > We did not see these errors until the tables grew over 80 Meg. > > In acct_history we are doing inserts only. In user_acct we are > > doing updates only. > > I estimate, at peak load, we are processing an average of five > transactions per second to user_acct. Spikes would probably go > as > high as 15-20 trans per second. On acct_history it would be more > like > one transaction per second. user_acct occurences outnumber > acct_history 5:1. > > Also notice the indices grew *after* reindexing and vacuuming. > We > did add 5000 cards today but aren't the indices suppposed to > update on > insert? > > We'd like to solve two problems here: > > 1. BTP_CHAIN errors cause system crash during peak traffic. > 2. Table sizes grow too much in short period of time. > > Best Regards, > > -Dave > > David Schanen : Atlas Telecom : mtv@scene.com > > ---------- Before Reindex and Vacuum --------- > > pgsql 176611328 May 6 02:13 acct_history > pgsql 55787520 May 6 02:13 acct_history_acct_no_idx > pgsql 120594432 May 6 02:17 user_acct > pgsql 22855680 May 6 02:17 user_acct_acct_no_idx > pgsql 31547392 May 6 02:17 user_acct_card_acct_sim_idx > pgsql 15908864 May 6 02:17 user_acct_card_no_idx > pgsql 9986048 May 6 02:17 user_acct_serial_no_idx > pgsql 12328960 May 6 02:17 user_acct_sim_idx > pgsql 8192 May 6 02:13 user_acct_state > pgsql 16384 May 2 03:00 user_acct_state_state_idx > > ---------- After Reindex and Vacuum --------- > > pgsql 176611328 May 6 02:13 acct_history > pgsql 55787520 May 6 02:13 acct_history_acct_no_idx > pgsql 81649664 May 6 02:41 user_acct > pgsql 29327360 May 6 02:41 user_acct_acct_no_idx > pgsql 45195264 May 6 02:41 user_acct_card_acct_sim_idx > pgsql 18595840 May 6 02:41 user_acct_card_no_idx > pgsql 15212544 May 6 02:41 user_acct_serial_no_idx > pgsql 12566528 May 6 02:41 user_acct_sim_idx > pgsql 8192 May 6 02:13 user_acct_state > pgsql 16384 May 2 03:00 user_acct_state_state_idx
В списке pgsql-hackers по дате отправления: