debian

You read into Documentation of lighttpd and discover the stat cache with fam which sounds really good. Save a lot of seek bandwidth of your disks by reducing the stat calls. What it doesnt tell you is that FAM is not even usable for moderate sizes of working sets. The number of monitorable directories is 1024 which is a limit of FAM as it uses select. This is so rediculus - lighttpd promoting FAM and fam beeing so poorly written (others would say - Its so '80s ...)

Link: FAM 1024 FD limitation

Whats the point in having dependencies when they are not used or maintained? With Debian/Sid you can install an MP3 player which cant play mp3's ... amarok: claims that xine cannot play mp3. Another bad example is pulseaudio - Why on earth cant i open padevchooser until i will install avahi which pulseaudio does not depend on. Seperating packages is a good thing but this makes them basically unusable. I switched to rhythmbox because of the whole amarok mess.

Debian - This is something Ubuntu solves a lot better.

"The most important part - Try to not break those machines" i said in #1 - Sir - We lost Nuernberg.

After reproducing it on my desk i found an interesting "grub" bug. All machines are currently equipped with lilo and i wanted to upgrade to grub. So - you issue an apt-get install grub then you need to replace the mbr and write a config. When you first issue an upgrade-grub you are boned. It will write a hd(0,0) as the root partition. When you grub-install /dev/hda first everything is fine and it'll detect the correct root partition.

While playing with the upgrade procedure i thought of an fast image based machine recovery. It seems partimage will only be able to cover single partitions not full disks including partioning.

One of the first problems to stumple upon is - How do i install/dist-upgrade without user interaction. The debconf part is easy - Just call apt-get dist-ugrade with DEBIAN_FRONTEND=noninteractive. But then you still get asked on whether you want the config files to get replaced which is dpkg's fault. There is a dpkg option to either take all new configs --force-confnew or to take all old config --force-confold. Now put everything together:

DEBIAN_FRONTEND=noninteractive \
apt-get \
-o Dpkg::Options::="--force-confnew" \
--force-yes \
-fuy \
dist-upgrade

If all packages would be nice this would be the solution. There are still packages which do everything on their own. Have a look at bind. It heavily complains even with this setup that /var/named exists and is non-empty. Another issue with bind is that the upgrade from bind to bind9 on etch tries to create a user bind. With my setup this fails because i am running libnss-ldap and the local machine may not create users in the ldap. So the first thing to do is to purge bind, disable ldap by changing nsswitch.conf and later to the reverse.

I am taking care on ~360 Debian machines at work. Those are still running Woody so i had a look on how to upgrade them automatically. Things to do would be:

• Upgrade woody -> sarge -> etch
• replace lilo by grub but take care on serial console
• replace handcraftet by debian kernel
• switch to udev
• repartition and resize2fs the last partition (/opt) to max disk
• switch from rdist to cfengine2 for config distribution

The most important part - Try to not break those machines as they are spread across germany in remote locations. The first thing i did was an inventory which machine had which Board, CPU, Disksize, Memsize and kernel type. I ran through all steps to upgrade a machine and manually noted the commands necessary. I'll report back which issues arise.