apt-get failures -- blah!

[Martin_J Carter]

Saith Chad Walstrom <chad at ima.umn.edu> 2001-11-09 17:32:00:

>I have an installation failure in rcS_fai that has been bugging me
for
>some time.  It happens when apt cannot find an install candidate for
a
>listed package.  Apt fails out and FAI continues on as if nothing
>happened.

Yup: We get similar effects here, for an HTTP-based FAI install,
whenever our local web cache system wimps out.  This invariably
happens for enormous files in the middle of a workstation install
during business hours.  Starting 45-min installations after 18:00
on a regular basis is not good for domestic bliss :-( .

[ snip ]

>To make the process a little easier, I changed the mounted disk check
to
>continue on if the mirror is already mounted instead of fail.  I then
>simply rerun rcS_fai from the shell on on the workstation.  Repeat
until
>I get a successful install.

Neat.  In our case, I put a Squid cache on the master node; not
only does it up the reliability of w/s installs, but it also brings
down
the install time from 45 mins to under 30 (once the files have been
requested once, and are safely in the cache).  Your approach looks
orthogonal to ours.

>What I'd like to know is what can we do to make apt-get failures less
>damaging to the install process.  How can we insert a bit more
>fault-tolerance to the process so that we can get a
"mostly-installed"
>system instead of a "base" system?  Do we need to patch apt to
provide
>the option to continue on in these cases?  Or, should we write a
helper
>app that breaks the package list into smaller useable chunks?  If we
run
>into a failure, it only affects those in packages in the smaller
chunk.

I did consider having each apt-get looped on until success, but that'd
only work with packages missing due to Web/FTP failure, and then only
in the absence of depencency breakages.  Why not optionally halt on
error? then we can see what's gone wrong as soon as possible, and
whether or not it can then be continued with a little manual
intervention.
Or have I missed something as usual?

>Thanks,
>Chad Walstrom <chewie at debian.org> 

That's *quite* enough from me for now :-) ,



        MaJoC
----
Dr M J Carter
Physics and Astronomy
University of Nottingham
E-mail: Martin.J.Carter at nottingham.ac.uk