Обсуждение: How to generate specific WAL records?
Hi, Does anyone know how to generate the following WAL records from psql? I'm now fixing pg_lesslog, which I reported a bug. Now code is almost okay and I'd like to add test tools to show pg_lesslog can handle all the WAL record correctly. XLOG_MULTIXACT_ZERO_OFF_PAGE XLOG_MULTIXACT_ZERO_MEM_PAGE XLOG_MULTIXACT_CREATE_ID XLOG_HEAP_LOCK XLOG_HEAP_INIT_PAGE XLOG_BTREE_INSERT_META XLOG_BTREE_DELETE_PAGE XLOG_BTREE_DELETE_PAGE_META XLOG_BTREE_DELETE_PAGE_HALF XLOG_GIN_CREATE_PTREE XLOG_GIN_SPLIT XLOG_GIN_VACUUM_PAGE XLOG_GIN_DELETE_PAGE XLOG_GIN_DELETE_LISTPAGE XLOG_GIST_PAGE_DELETE So far, I'm using conventional BTREE, btree_gin and btree_gist for the test, as well as 2PC and savepoint. Any information is welcome. Thank you very much in advance; ------ Koichi Suzuki
Koichi Suzuki escribió: > Hi, > > Does anyone know how to generate the following WAL records from psql? > > I'm now fixing pg_lesslog, which I reported a bug. Now code is > almost okay and I'd like to add test tools to show pg_lesslog can > handle all the WAL record correctly. > > XLOG_MULTIXACT_ZERO_OFF_PAGE > XLOG_MULTIXACT_ZERO_MEM_PAGE > XLOG_MULTIXACT_CREATE_ID For these, you need to do SELECT FOR SHARE of a tuple in more than one concurrent session. I think HEAP_LOCK is SELECT FOR UPDATE. > XLOG_BTREE_DELETE_PAGE > XLOG_BTREE_DELETE_PAGE_META > XLOG_BTREE_DELETE_PAGE_HALF These are caused during vacuum of a btree. Create an index, populate it, then remove all items. Then vacuum twice. -- Alvaro Herrera http://www.CommandPrompt.com/ PostgreSQL Replication, Consulting, Custom Development, 24x7 support
Alvaro-san; Thank you for a great advice. I successfully generated all the WAL records listed below. By deleteing and VACUUMing a table with btree_gist based index, I was successful to generate all the GIST WAL records. Still have the following WAL record to create: CLOG_TRUNCATE XLOG_HEAP_INIT_PAGE XLOG_BTREE_INSERT_META XLOG_GIN_CREATE_PTREE XLOG_GIN_DELETE_PAGE I have to run good amount of transactions before VACUUM FREEZE to create CLOG_TRUNCATE. I'll try. It's very helpful if anybody knows how to generate XLOG_HEAP_INIT_PAGE and XLOG_BTREE_INSERT_META. I'll ask Oleg-san how to generate GIN-related WALs. Thank you very much; ---------- Koichi Suzuki 2010/4/14 Alvaro Herrera <alvherre@commandprompt.com>: > Koichi Suzuki escribió: >> Hi, >> >> Does anyone know how to generate the following WAL records from psql? >> >> I'm now fixing pg_lesslog, which I reported a bug. Now code is >> almost okay and I'd like to add test tools to show pg_lesslog can >> handle all the WAL record correctly. >> >> XLOG_MULTIXACT_ZERO_OFF_PAGE >> XLOG_MULTIXACT_ZERO_MEM_PAGE >> XLOG_MULTIXACT_CREATE_ID > > For these, you need to do SELECT FOR SHARE of a tuple in more than one > concurrent session. > > I think HEAP_LOCK is SELECT FOR UPDATE. > > >> XLOG_BTREE_DELETE_PAGE >> XLOG_BTREE_DELETE_PAGE_META >> XLOG_BTREE_DELETE_PAGE_HALF > > These are caused during vacuum of a btree. Create an index, populate > it, then remove all items. Then vacuum twice. > > -- > Alvaro Herrera http://www.CommandPrompt.com/ > PostgreSQL Replication, Consulting, Custom Development, 24x7 support >
On Wed, 2010-04-14 at 11:13 +0900, Koichi Suzuki wrote: > Thank you for a great advice. I successfully generated all the WAL > records listed below. By deleteing and VACUUMing a table with > btree_gist based index, I was successful to generate all the GIST WAL > records. It would be a very useful test case to publish. There are already a number of standby related tests, this would be very useful for general case testing, not just for pg_lesslog. -- Simon Riggs www.2ndQuadrant.com
Thanks for encouraging comment. I'm still struggling to generate remaing WAL records. ---------- Koichi Suzuki 2010/4/14 Simon Riggs <simon@2ndquadrant.com>: > On Wed, 2010-04-14 at 11:13 +0900, Koichi Suzuki wrote: > >> Thank you for a great advice. I successfully generated all the WAL >> records listed below. By deleteing and VACUUMing a table with >> btree_gist based index, I was successful to generate all the GIST WAL >> records. > > It would be a very useful test case to publish. There are already a > number of standby related tests, this would be very useful for general > case testing, not just for pg_lesslog. > > -- > Simon Riggs www.2ndQuadrant.com > >
Koichi Suzuki <koichi.szk@gmail.com> wrote: > 2010/4/14 Simon Riggs <simon@2ndquadrant.com>: >> It would be a very useful test case to publish. > I'm still struggling to generate remaing WAL records. Sure, but when you've got it all, please share. I'd like to see us have a much larger set of tests than the "make check" regression tests which would get run less frequently but test a lot more. This sounds like a good one to include. -Kevin
Kevin Grittner wrote: > Koichi Suzuki <koichi.szk@gmail.com> wrote: > > 2010/4/14 Simon Riggs <simon@2ndquadrant.com>: > > >> It would be a very useful test case to publish. > > > I'm still struggling to generate remaing WAL records. > > Sure, but when you've got it all, please share. I'd like to see us > have a much larger set of tests than the "make check" regression > tests which would get run less frequently but test a lot more. This > sounds like a good one to include. Agreed. This test belongs in our CVS tree so we can maintain it as we add new WAL types. -- Bruce Momjian <bruce@momjian.us> http://momjian.us EnterpriseDB http://enterprisedb.com
Now I've tested almost all WAL record, including CLOG_TRUNCATE and all the GIST and GIN-related WALs. I'm still struggling to find how to have XLOG_HEAP_INIT_PAGE and XLOG_BTREE_INSERT_META. I'm trying to find how to have these WAL records but it's a great help if anyone suggests me how. Thank you very much in advance; ---------- Koichi Suzuki 2010/4/15 Bruce Momjian <bruce@momjian.us>: > Kevin Grittner wrote: >> Koichi Suzuki <koichi.szk@gmail.com> wrote: >> > 2010/4/14 Simon Riggs <simon@2ndquadrant.com>: >> >> >> It would be a very useful test case to publish. >> >> > I'm still struggling to generate remaing WAL records. >> >> Sure, but when you've got it all, please share. I'd like to see us >> have a much larger set of tests than the "make check" regression >> tests which would get run less frequently but test a lot more. This >> sounds like a good one to include. > > Agreed. This test belongs in our CVS tree so we can maintain it as we > add new WAL types. > > -- > Bruce Momjian <bruce@momjian.us> http://momjian.us > EnterpriseDB http://enterprisedb.com >