Tales from the Server Room

I've always thought it's better to learn from someone else's mistakes than from my own. In this column, Kyle Rankin or Bill Childers will tell a story from their years as systems administrators while the other will chime in from time to time. It's a win-win: you get to learn from our experiences, and we get to make snide comments to each other. Kyle tells the first story in this series.

I was pretty excited about my first trip to the London data center. I had been to London before on vacation, but this was the first time I would visit our colocation facility on business. What's more, it was the first remote data-center trip I was to take by myself. Because I still was relatively new to the company and the junior-most sysadmin at the time, this was the perfect opportunity to prove that I knew what I was doing and could be trusted for future trips.

The maintenance was relatively straightforward. A few machines needed a fresh Linux install, plus I would troubleshoot an unresponsive server, audit our serial console connections, and do a few other odds and ends. We estimated it was a two-day job, but just in case, we added an extra provisional day.

[Bill: If I remember right, I had to fight to get that extra day tacked onto the trip for you. We'd learned from past experience that nothing at that place seemed easy at face value.]

Even with an extra day, I wanted this trip to go smoothly, so I came up with a comprehensive plan. Each task was ordered by its priority along with detailed lists of the various commands and procedures I would use to accomplish each task. I even set up an itemized checklist of everything I needed to take with me.

[Bill: I remember thinking that you were taking it way too seriously—after all, it was just a kickstart of a few new machines. What could possibly go wrong? In hindsight, I'm glad you made all those lists.]

The first day I arrived at the data center, I knew exactly what I needed to do. Once I got my badge and was escorted through multiple levels of security to our colocation cages, I would kickstart each of the servers on my list one by one and perform all the manual configuration steps they needed. If I had time, I could finish the rest of the maintenance; otherwise, I'd leave any other tasks for the next day.

Now, it's worth noting that at this time we didn't have a sophisticated kickstart system in place nor did we have advanced lights-out management—just a serial console and a remotely controlled power system. Although our data center did have a kickstart server with a package repository, we still had to connect each server to a monitor and keyboard, boot from an install CD and manually type in the URL to the kickstart file.

[Bill: I think this experience is what started us down the path of a lights-out management solution. I remember pitching it to the executives as “administering from the Bahamas”, and relaying this story to them was one of the key reasons that pitch was successful.]

After I had connected everything to the first server, I inserted the CD, booted the system and typed in my kickstart URL according to my detailed plans. Immediately I saw the kernel load, and the kickstart process was under way. Wow, if everything keeps going this way, I might even get this done early, I thought. Before I could start making plans for my extra days in London though, I saw the kickstart red screen of death. The kickstart logs showed that for some reason, it wasn't able to retrieve some of the files it needed from the kickstart server.

Great, now I needed to troubleshoot a broken kickstart server. Luckily, I had brought my laptop with me, and the troubleshooting was straightforward. I connected my laptop to the network, eventually got a DHCP lease, pointed the browser to the kickstart server, and sure enough, I was able to see my kickstart configuration files and browse through my package repository with no problems.

I wasn't exactly sure what was wrong, but I chalked it up to a momentary blip and decided to try again. This time, the kickstart failed, but at a different point in the install. I tried a third time, and it failed at the original point in the install. I repeated the kickstart process multiple times, trying to see some sort of pattern, but all I saw was the kickstart fail at a few different times.

The most maddening thing about this problem was the inconsistency. What's worse, even though I had more days to work on this, the kickstart of this first server was the most important task to get done immediately. In a few hours, I would have a team of people waiting on the server so they could set it up as a database system.

Here I was, thousands of miles away from home, breathing in the warm exhaust from a rack full of servers, trying to bring a stubborn server back to life. I wasn't completely without options just yet. I had a hunch the problem was related to DHCP, so I pored through the logs on my DHCP server and confirmed that, yes, I could see leases being granted to the server, and, yes, there were ample spare leases to hand out. I even restarted the DHCP service for good measure.

Finally, I decided to watch the DHCP logs during a kickstart. I would start the kickstart process, see the machine gets its lease, either the first time or when I told it to retry, then fail later on in the install. I had a log full of successful DHCP requests with no explanation of why it didn't work. Then I had my first real clue: during one of the kickstarts, I noticed that the server had actually requested a DHCP lease multiple times.

Even with this clue, I started running out of explanations. The DHCP server seemed to be healthy. After all, my laptop was able to use it just fine, and I had a log file full of successful DHCP requests. Here I turned to the next phase of troubleshooting: the guessing game. I swapped cables, changed what NIC was connected and even changed the switch port. After all of that, I still had the same issue. I had kickstarted the machine so many times now, I had the entire list of arguments memorized. I was running out of options, patience and most important, time.

[Bill: I remember seeing an e-mail or two about this. I was comfortably ensconced at the corporate HQ in California, and you were working on this while I was asleep. I'm sure I'd have been able to help more if I'd been awake. I'm glad you were on the case though.]

I was now at the next phase of troubleshooting: prayer. Somewhere around this time, I had my big breakthrough. While I was swapping all the cables around, I noticed something interesting on the switch—the LEDs for the port I was using went amber when I first plugged in the cable, and it took quite a bit of time to turn green. I noticed that the same thing happened when I kickstarted my machine and again later on during the install. It looked as though every time the server brought up its network interface, it would cause the switch to reset the port. When I watched this carefully, I saw during one install that the server errored out of the install while the port was still amber and just before it turned green!

What did all of this mean? Although it was true that the DHCP server was functioning correctly, DHCP requests themselves typically have a 30-second timeout before they give an error. It turned out that this switch was just hovering on the 30-second limit to bring a port up. When it was below 30 seconds I would get a lease; when it wasn't, I wouldn't. Even though I found the cause of the problem, it didn't do me much good. Because the installer appeared to reset its port at least three times, there was just about no way I was going to be able to be lucky enough to get three consecutive sub-30-second port resets. I had to figure out another way, yet I didn't manage the networking gear, and the people who did wouldn't be awake for hours (see sidebar).

[Bill: One of the guys I worked with right out of college always told me “Start your troubleshooting with the cabling.” When troubleshooting networking issues, it's easy to forget about things that can affect the link-layer, so I check those as part of the cabling now. It doesn't take long and can save tons of time.]

I started reviewing my options. I needed some way to take the switch out of the equation. In all of my planning for this trip, I happened to bring quite a toolkit of MacGyver sysadmin gear, including a short handmade crossover cable and a coupler. I needed to keep the original kickstart server on the network, but I realized if I could clone all of the kickstart configurations, DHCP settings and package repositories to my laptop, I could connect to the machine with a crossover cable and complete the kickstart that way.

After a few apt-gets, rsyncs, and some tweaking and tuning on the server room floor, I had my Frankenstein kickstart server ready to go. Like I had hoped, the kickstart completed without a hitch. I was then able to repeat the same task on the other two servers in no time and was relieved to send the e-mail to the rest of the team saying that all of their servers were ready for them, right on schedule. On the next day of the trip, I was able to knock out all of my tasks early so I could spend the final provisional day sightseeing around London. It all goes to show that although a good plan is important, you also should be flexible for when something inevitably goes outside your plan.

[Bill: I'm glad you planned like you did, but it also highlights how important being observant and having a good troubleshooting methodology are. Although you were able to duct-tape a new kickstart server out of your laptop, you could have spent infinitely longer chasing the issue. It's just as important to know when to stop chasing a problem and put a band-aid in place as it is to fix the problem in the first place.]