I am mostly responsible for the scripts that upgrade our software from one version to the next. For the most part, it’s pretty straight forward, since we use apt-rpm to take care of upgrading rpms and making sure their dependencies are fulfilled. There is a bit of a hitch in that some of our rpms were made by people who don’t really get rpm, so the rpm just installs a tar file and the %post script unpacks the tar. Trust me, that’s more bizarre than you think, even if you think it’s pretty bizarre. Because of that and because of dependencies, we can’t just do a “dist-upgrade”, but have to upgrade our rpms one at a time.
But the biggest twist on the upgrade is that going from 3.3 to 3.6 of our software involves going from RedHat 7.3 to Centos 3.4. (Don’t ask what happened to versions 3.4 and 3.5 of our software – it’s too painful.) I am not quite proud of the horrible hack I put together to do that, but it took a bunch of work to get it working so that they can just put a DVD in the drive and type /mnt/cdrom/upgrade now and it mostly works. It uses grub on the hard disk to boot the DVD with a custom kickstart file, formats all the partitions but one, installs the new CentOS, and then uses a custom finish script to reinstall our software and restore the backed up configuration off that one partition that wasn’t formatted.
Actually, as an aside, all of the upgrades only “mostly work”. Partly that’s because we were stupid enough to put Dell computers on the customer sites, so it’s hit or miss whether they’ll even reboot after a year or two’s continuous use. But mostly it’s because rpm has an intermittent bug in the locking code that RedHat was told about at least 3 years ago, and still hasn’t fixed. Which means that sometimes apt-get fires up rpm -U, and rpm just hangs. And of course I get blamed because the upgrade doesn’t entirely work, although I think they’re starting to realize it’s not my fault.
When the upgrades don’t work, I usually get dragged in by the customer support people to log into the customer site to fix it. And usually it’s pretty straight forward – reboot the machine with the locked rpm database, and then manually step through the steps the upgrade script would have taken on that machine. Although the customer support people have recently learned that they can just re-run the whole upgrade script, because the apt-get install portion won’t do anything if it was already done.
Today I got dragged in because of a bigger problem – a customer site that is running version 6.2 of our software had a hardware crash on the main server. They sent the customer a new server, but for some stupid reason the fulfillment house that sent out the new server had version 3.3 of our software installed on it. I guess customer support sent them a 3.3->3.6 upgrade DVD, and tried to upgrade it to 3.6 remotely, and then upgrade from 3.6 to 6.2. But at that point they noticed nothing was working.
I investigated, and discovered that most of the rpms weren’t installed. Also, nothing was backed up properly in the partition that doesn’t get reformatted, so it hadn’t been restored correctly. So I decided to go back to the 3.6 version to see if I could get it working there and then go forward. Fortunately they’d left the 3.3->3.6 upgrade DVD in the drive. I ran the /mnt/cdrom/upgrade script and came back a few hours later. And sure enough, only 3 of the 8 rpms that are normally installed were installed. I tried to manually install them, but the first one failed because it had a dependency on the mozilla rpms, and the mozilla rpm was corrupt. It didn’t matter whether I tried the one in the apt repository, or on the DVD itself. It was hosed.
At this point, I gave up and said that they’d have to ship out another replacement unit with the proper version of our software installed. So much for being the hero.
But as I was leaving work, I had a few ideas on how I could have repaired that mozilla rpm and gotten it working. But I was too late. I figured even if I went back everybody was gone as well, and besides they’d already ordered the replacment – my reputation as a hero is ruined. Sigh.
Experience on the Fedora lists suggests that it might not be a RPM bug, but a kernel bug that has been fixed during the 2.6.19 timeframe.
This bug may or may not be the culprit, but was found at about the right time, affects Berkeley databases, and is plausible.
Since fixed Fedora kernels came out, the only complaints about RPM database corruption have come from people who were running old kernels.
I don’t know if this could help you…
James, the bug you reference wasn’t introduced until 2.6.18. We’ve been seeing rpm lockups since we started with 2.4.x in RedHat 7.3.
The bug I’m referring to is, if I remember correctly, an in-memory locking system that IBM wrote that if it deadlocks there is no way to fix short of rebooting.
In my experience, you need an occasional failure to help manage expectations (others, not yours). Once you get a reputation for pulling people’s chestnuts out of the fire, they will expect it from you. Not only that, but they will be less careful, since they know you will be there to save them when they screw up (again).