Re: [Condor-users] auto-clustering of attributes

["Matt Hope" <matthew.hope@xxxxxxxxx>]

On 6/26/07, Todd Tannenbaum <tannenba@xxxxxxxxxxx> wrote:
> Horvátth Szabolcs wrote:
> > Hi,
> >
> > The auto-clustering of attributes seems to be negating the effect of
> > changing a custom attribute that is used in the starter's rank expression.
> > I'd like to modify the importance of a job after submission but it has
> > already been clustered and attribute changes have no effect.
>
> Right after you change the attribute (via condor_qedit I am guessing),
> set the job attribute "AutoClusterId" to either -1 or UNDEFINED.  That
> will cause Condor to recompute the auto cluster for that job.
>
> Although I would be surprised if the above is really what is happening.
>   If any attribute listed in the job ad attribute "AutoClusterAttrs"
> (this lists all the job attributes considered to be significant) changes
> value, then Condor automatically clears out the AutoClusterId attribute
> so it gets recomputed at the next negotiation cycle.  So the above
> problem "shouldn't be happening" (TM).  ;)

It is possible the OP was referring to the RANK *after* the job has
begun running.

This appears *not* to be recomputed if the underlying values driving
it or the calculation mechanism is changed. At least on 6.6.

This means massive changes to the way RANK is calculated such that
previous numbers are wholly inappropriate in relation to the old ones
can cause your farm to suddenly toss all of it's jobs away. (Happened
to me once - never allow RANK expression changes to be rolled out to
your pool en-masse. especially if the new numbers are going to be
inflated).

I do not know if this is a feature of the startd, the negotiator or
interaction of both (at the time I didn't have the source so just took
it as a feature and left it at that) but I believe recalculation is
the better of the options as it has the more 'obvious' behaviour.

So in answer to the OP - at the moment what you appear to want is not
possible as far as I can tell - I have not exhausted the possibilities
available from the periodic_* mechanism but I have a workaround (not
changing it except when the farm needs a reboot :).

Matt