On Fri, Mar 18, 2005 at 03:00:47PM +1100, Lars Norved wrote: > Lothar M?rkle <[EMAIL PROTECTED]> writes: > >>> Assuming you have a cgi-like application with many processes > >>> that just looks up a row, displays and then exits. You can > >>> simple use the rename call to atomically replace the db file > >>> (see man 2 rename for bordercases) with another and without > >>> locking. > >>> > >>> lothar > >> > >> but this will lose any update currently begin written by any > >> of the active cgi processes - even if they are in the middle > >> of a transaction? > > > > yes, this will only work with a read-only db, e.g. all your data > > is pre-calculated once a day to a second db file, and then replace > > the primary read-only db with it. > > > > But with read-only, one should avoid transactions at all. There > > will be the risk of a -journal that doesn't correspond to the > > db-file. I think this will break if rolled back? > > > > lothar > > I need to copy a 'live' master database file and overwrite a read > only file for pseudo replication. > > How do I ensure the database is flushed and acquire an exclusive > lock on the master database? > > Will something like the code below do it? > > sqlite3_exec(db, "BEGIN EXCLUSIVE TRANSACTION", NULL, NULL, error); > copyFiles(); > sqlite3_exec(db, "COMMIT", NULL, NULL, error);
i think you have to use sqlite3_close to be shure all data is written. then copy. > Also are you sure it is completely safe to replace a read only > database while reads are currently being done? I'm using this with cdb, which is a read-only constant-dbm like thingy with put and get only(cr.yp.to/cdb.html) on unix, this is save for read only one shot processes reading file a and one updating process. fd = open( a.tmp) print( fd, "some new stuff") close( fd ) rename( a.tmp, a) no process reading file a will see a broken file and no locking is used, because rename is atomic as long you stay in the same filesystem. If you have only readers, i see no reason why an app using sqlite should break. But the readers have to reopen the db. lothar
