Re: serverlog rotation/functions

Tom Lane wrote:
> Andreas Pflug <pgadmin(at)pse-consulting(dot)de> writes:
>
>>In-use check is easy for the backend, if the syslog process publishes
>>the current logfile's timestamp in sharedmem.
>
>
> You really haven't absorbed any of the objections I've raised, have you?
> I don't want the log process connected to shared mem at *all*, and see
> no particularly good reason why it should be.

Why shouldn't the process announce the logfile timestamp and its pid
*writeonly*, so other backends know about it?

At least the pid must be distributed like this, just as bgwriter does.
I understand perfectly that postmaster and logger are very critical
processes, so they should be dependent on as few resources as possible.
The logger works without shmem in general, but how to reach it if its
pid is unknown?

> The directory should definitely be a GUC variable. The individual
> filenames should probably be of the form <prefix><timestamp>, where
> the server dictates the format of the timestamp (and we choose it so
> that the names sort correctly). We could let the prefix be
> user-selectable or make it hard-wired; I don't have a strong feeling
> about that either way.

Agreed.

Regard,
Andreas