[Bizgres-general] Another use case

[Jim Nasby]

I'm not sure if this fits exactly into queuing or not, but lowering  
the priority of a specific backend is something that comes up every  
few months.

Begin forwarded message:

> From: "Aaron Bono" <postgresql at aranya.com>
> Date: August 1, 2006 2:52:32 PM CDT
> To: "Scott Marlowe" <smarlowe at g2switchworks.com>
> Cc: "Thomas Pundt" <thomas.pundt at rp-online.de>, pgsql- 
> admin at postgresql.org
> Subject: Re: [ADMIN] Controlling CPU Usage in PostgreSQL
>
> On 8/1/06, Scott Marlowe <smarlowe at g2switchworks.com> wrote:
> On Tue, 2006-08-01 at 13:59, Aaron Bono wrote:
> > On 7/31/06, Thomas Pundt <mlists at rp-online.de> wrote:
> >         On Monday 31 July 2006 10:06, Aaron Bono wrote:
> >         | Is there a way I can tell PostgreSQL to give session a low
> >         priority so even
> >         | if it does take the full CPU, it only does so if it would
> >         otherwise be
> >         | idle?
> >
> >         the "nice" command might do what you want...
> >
> >
> > OK, so I tried:
> >
> > su - postgres -c "nice -n 19 psql my_db"
> >
> > The problem is, the psql command has a nice value of 19 but the
> > PostgreSQL server process that psql has connected to is running  
> with a
> > nice value of 0.  My assumption is that, if I then run my functions,
> > psql will get low priority but it is the server process that is
> > running with normal priority that will still use up all the CPU.
> >
> > Bottom line, I am skeptical if this will really achieve my goal - to
> > have the functions run with low priority.
> >
> > Is there a way to tell PostgreSQL to change the nice value of a
> > particular connection?  I guess I could use renice but that means a
> > lot of shell scripting to determine what PID to change - I am not a
> > shell script expert and would rather not pursue that option.
>
> This is one of those things that only seems like a good idea at the
> time.
>
> While nice-ing the whole of a postgresql server is doable, nice-ing
> individual connections is no advisable.
>
> You can reach a bad state caused by "priority inversion".  Search the
> lists (not just admin) for that phrase and you should find out what I
> mean.  This should probably be a FAQ, as it shows up every few months.
>
> I am guessing my best bet is to wait until performance is becomming  
> an issue (or just before) and then create a data warehouse on a  
> separate server.  That would be much easier to implement and would  
> isolate the reporting from production.
>
> Thanks for the input.
>
> ==================================================================
>    Aaron Bono
>    Aranya Software Technologies, Inc.
>    http://www.aranya.com
> ==================================================================