On Feb 23, 2021, at 11:41 AM, Richard Damon <rich...@damon-family.org> wrote:
> On 2/23/21 2:18 PM, Wietse Venema wrote: >> Ron Garret: >>>> If we take this route, then there needs to be a new field in the >>>> Postfix sqlite config file that controls the time limit. >>> Not necessarily. You could just hard-code a reasonable value (like >>> 1 second), or make it a #define so you need a recompile to change >>> it. That?s sub-optimal, obviously, but still a major improvement >>> over the current situation for very little effort and no down-side >>> that I can see. >> The limit should be configurable. It takes: >> >> - one line of code to define a C variable, >> >> - one line of code to read its value from an sqlite_table configuration >> file (or to use a documented default value), >> >> - a few lines of text to document the new field in the sqlite_table manpage. >> >> Wietse > > One thng to look at is WAL mode. WAL mode increases the cost of writes > to the database, as all writes become two stage, first to the WAL > journal, and then flushed to the main database (called A checkpoint), > and reads reads can get a bit more expensive if the second stage of the > write gets delayed by long accesses (but that may not be an issue with > postfix). > > In exchange, the database allows for simultaneous reads and writes, > except possibly for the period when the second phase of the writes are > occurring, and it will try to allow as much overlap there as possible, > and try to find a time when no readers are active to do this operation. > > Without a busy timeout being set, the reader should only get a busy in > fairly rare conditions, the main one being if the last connection to the > database is closing, then SQLite does some cleanup that locks the > database for just a little bit, or if the last connection 'crashes' than > the next connection will do some cleanup. Even a fairly short busy wait > should handle these cases most of the time. WAL mode was previously discussed here: https://marc.info/?l=postfix-users&m=160096626120296&w=2 The upshot appears to be this, at least as things currently stand: > DO NOT use SQLite as a Postfix backend database updated live while > Postfix is running. rg
