El día 22 de agosto de 2011 18:39, Greg Smith <greg@xxxxxxxxxxxxxxx> escribió: > On 08/22/2011 05:07 PM, Martín Marqués wrote: >> >> My question regarding your answer is, why is it important for the >> first page after a checkpoint and not on other page writes? >> > > The first time a page is written after a checkpoint, when full_page_writes > is on, the entire 8K page is written out to disk at that point. The idea is > that if the page is corrupted in any way by a partial write, you can restore > it to a known good state again by using this version. After that copy, > though, additional modifications to the page only need to save the delta of > what changed, at the row level. If there's a crash, during recovery the > full page image will be written, then the series of deltas, ending up with > the same data as was intended. > > This whole mechanism resets again each time a checkpoint finishes, and the > full page writes start all over again. One of the main purposes of > checkpoints are to move forward the pointer of how far back crash recovery > needs to replay from. Starting each new checkpoint over again, with a full > copy of all the data modified going into the WAL, it is part of that logic. Still something missing (for me :-)): Checkpoint happens, logs get flushed to disk, WAL segments get archived if archive is on, and then the are recycled. Now a new transaction get commited, and at least 1 page has to go to WAL (those are 8Kb). full_page_writes garantees all 8Kb are written or nothing. After this thansaction comes another, but here you don't need the garantee of all 8Kb in WAL (you say because deltas are enough). Why aren't deltas good enough for the first 8Kb? Is there other information in the first 8Kb that make those more important? -- Martín Marqués select 'martin.marques' || '@' || 'gmail.com' DBA, Programador, Administrador -- Sent via pgsql-general mailing list (pgsql-general@xxxxxxxxxxxxxx) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-general
