Alvaro Herrera wrote:
Andrew Dunstan wrote:
The attached patch makes a very small but useful change to the behaviour of
log_line_prefix, by enabling the start time (%s) and cookie (%c) logging to
occur for all backends rather than just for session processes (i.e.
backends started for a client connection). We actually need almost all of
this patch, with or without the change in behaviour, so we can put the
cookie in CSVlogs (which I'm still working on), since the cookie+line
number make the natural primary key for the logs. The actual change in
behaviour from this patch comes from the removal of 2 "if (MyProcPort)"
lines in elog.c. Given that, can I sneak this in or should I wait for 8.4
given we're long past feature freeze?
Thinking again about the feature itself I wonder if it actually makes
sense -- maybe it does make sense to be able to display the session ID,
but the start time? Why would anyone care about the start time of
syslogger or bgwriter? We don't even have a use for the "hey, this
process was started" log line, why would anyone care about having the
start time in the log line prefix?
Actually having the cookie in all processes is another matter, as far as
it's useful for CSV logs. But then, is it? Maybe the auxiliary
processes should identify themselves with fixed cookies or something
particular that lets one distinguish, say, a bgwriter from a syslogger,
but is there a case from distinguishing one bgwriter from another?
It's not about distinguishing one bgwriter from another, it's about
distinguishing it from any other process at any time whatsoever that has
had the same pid. cookie+linenumber should be unique. pid+linenumber
isn't. (And every process gets its own line number sequence, so we can't
just give, say, all the bgwriter processes the same cookie). Logging the
start time on its own isn't much extra benefit, although I expect log
parsers will find it nicer to be able to handle a more consistent
logging style rather than having to handle non-session processes as a
special case. But having the cookie available in all cases is the whole
point of this - I wouldn't have done it unless I had needed to be able
to set a primary key for loadable logs.
If you want to invent some other style of cookie we can look at that.
Back when we looked at it originally nobody came up with a better
suggestion than process_start.pid. But that surely would be for a later
So, short answer - yes, I think it makes sense. But if there's any
serious argument I won't change the observable behaviour in elog.c, just
---------------------------(end of broadcast)---------------------------
TIP 9: In versions below 8.0, the planner will ignore your desire to
choose an index scan if your joining column's datatypes do not