Life is funny sometimes when everything is rolling along according to plan; it seems like that’s when one should most be on their toes. It’s almost expected to think that during the week of my year anniversary at SourceClear, the largest outage (complete with two days of downtime over multiple teams) of my career would hit. What’s the best part of this story? The outage was caused in no small part, by yours truly. So for the sake of documentation, and transparency, follow along with me as we explore what I did wrong to break everything, and what we did right to fix it.
So to start, this was an out in our QA/Dev Kubernetes cluster, not production cluster. So, not production but when the environments you’re in charge of has been pretty stable under your watch, any downtime is downtime. Any lost productivity is a bummer, especially when you’re the sole cause of the issues. A lack of attention to detail while planning upgrades, and unfamiliarity with tooling. No excuses, this was 100% avoidable.
The back story is that we have been preparing for an AWS region migration, and I have been spinning up (and then tearing down) multiple Kubernetes clusters daily, using
kubespray, to ensure I know what to do when it doesn’t work, and document all steps involved in the process.
During one of these cycles, I accidentally deleted some IAM roles (but I did re-add them!) with Terraform. Turns out, IAM roles and instance profiles for EC2 instances? Kubernetes needs those for the
cloud-controller-manager! I re-added the roles with the proper permissions based on the
kubespray documentation, and nothing looked amiss.
The signs of trouble we first saw (which later came back around a week later) were errors in the
journalctl logs from
kubelet. Since the cluster was provisioned with
kubespray, the team and I opted to rerun
kubespray. This should detect any anomalous settings and such in our cluster, and make those changes back to what is defined in the playbook. Turns out, that lead to mistake two.
I’m not quite sure why I thought that a
kubespray run would fix things. It doesn’t lay anything down for IAM roles (that’s terraform), but I guess at this point I was grasping at straws? We had been seeing some intermittent issues with kubelet since the week of Thanksgiving. I attempt to re-run
kubespray and it fails once or twice but I eventually had a successful run, and to no one’s surprise, it didn’t fix anything but it made things worse…
It was at that point I realized, that
kubespray had run successfully, with the default variables, not my specific environment variables. So at this point, everything not in the
kube-system namespace was falling over. But wait, it gets better, I hadn’t known this was possible, but looking back on it, I see no reason why it shouldn’t be possible, but due to a misplaced variable in my vars file…I now had two copies of all the control plane pods. Two copies of the
kube-scheduler, and two CNI Plugins running at the same time.
So at this point, I was looking around for other things to try and get me back in a steady state, but it was rather late in the evening, we’d hit 13 hours in the day. It was time to call it a night, our QA outage would roll into day 2. This is especially painful when you have a team based on the other side of the world that was looking to use this cluster during their day, I had wasted a US-based teams day, and would now waste another’s.
The next day, with a good night’s sleep, I learn by reading the Ansible docs, that the
group_vars are based on the relative location of the Ansible inventory file. Well, I should mention that I had checked in my
inventory files to another git repo, that I was referencing by absolute path. And that was where Ansible/Kubespray was looking for environment variables. On a hunch, before the re-run of
kubespray, I checked my IAM roles. And as it turns out, I had missed a permission for the role. It didn’t match what we had in another account’s cluster. Adding that permission immediately resolved the issue with the
kubelet logging its inability to auth to the cloud provider. Now as a reminder, that was the initial issue, and by trying to fix that, I had made things worse.
Let’s get low hanging fruit first. Scale the second CNI plugin to 0 pods, and fix the multiple control plane pods. Back in business. Well halfway in business. We were so close to being in a good spot. Our cluster was up and accessible, we could make API calls, and pods were scheduling… but there were endless
CrashLoopBackOff for all new pods. WHY?
At this point, the team started looking more into DNS. If the CNI is correct, then maybe it’s a
KubeDNS issue? We tried a rolling restart on all those and see if anything clears up but to no avail. I noticed at this point that
KubeDNS was removed from Kubernetes 1.13, and while we weren’t running 1.13 yet, maybe this was the right time to tear out
KubeDNS and add
CoreDNS? Well, that’s just what I did. Now you may think this wasn’t a mistake, and it truly wasn’t, but it was a decent chunk of wasted time as
KubeDNS was in no way causing issues, but we did get the current Kubernetes DNS when all was said and done, something that would have had to happen eventually.
This is when we hit a turning point in our troubleshooting process. It turns out, another one of those
kubespray default values was a pod CIDR. Our desired CNI plugin was sitting in the right subnet, but pods were still getting the wrong CIDR, but from where? I have removed the secondary plugin.
I started going down my list of things I know to check. Stop/start
kubelet. Bounce the CNI networking pods. What about the Docker network? Nothing. No luck. All new pods have the incorrect (
kubespray default) CIDR. But where are they getting that value from?! There aren’t any rogue CNI bits left in the cluster, right? No pods, nothing!
Turns out that secondary CNI pods had left a config file in the /etc/cni/net.d/ directory. This is the directory that our chosen CNI reads from (populated by ConfigMap) but for some reason, that file
(10-<secondary-plugin>.conf) was causing issues by coexisting in that directory with
So we NUKED those secondary configs from orbit, reboot all our nodes…PODS! WITH PROPER CIDRs! Miraculous!
So all we have to do is go through all the nodes and delete that file and reboot? I know how to do all those things! (Side Note: We could have most likely done less than a full reboot, but in the interest of being thorough, we did one anyway).
I began moving through the list of nodes. ssh into the node, rm the config file in question, and sudo reboot. Piece of cake. I get a message from another engineer that I had missed a node. We have a decent number so that is entirely possible. They’re already ssh-ed in so they handled it. Then they found another I missed, and another.
I was offended. How could they assume me of missing more than half? I asked what commands they ran, are they rebooting through the console or terminal? It all lined up! What did I do wrong? So I asked…
“What file are you deleting?”
(10-<secondary-plugin>.conf), what did you delete?”
“Oh, I deleted
Why did I delete the config file from the CHI Plugin that we wanted you might ask? Well, I have a valid reason, that sounded more intelligent before I typed my reply. I thought that perhaps that config was corrupted in some way, and deleting it would load a new config from the ConfigMap. I promise it sounded good in my head, but I needed to delete
(10-<secondary-plugin>.conf), as that was causing the issues. Well now I know, and believe me, this other engineer with never let me live this down. We deleted the proper file (as I had deleted the wrong file on all the nodes) and everything came back up to where we wanted to be. Service restored, and it only took two days.
This was started by a mistake. A medium sized mistake, but a mistake I made nonetheless. Two days of downtime cascaded from missing a detail in my process. I then followed that up with 3-4 other mistakes that got us no closer to fixing the issue.
I am proud of my troubleshooting process for the most part. I think while in the thick of things, I was asking the right questions and looking in the right places. By being unaware the Instance Profile issue persisted, all my other troubleshooting wasn’t helping, and may have made things worse.
But, I took a lot of notes, and I learned about some parts of the infrastructure that I hadn’t had a lot of exposure to. I also learned a lot about CoreDNS and CNI Plugins and got to configure both, even though the DNS was never the issue, and the CNI Plugins were “too configured” if that is possible.
It wasn’t production downtime but it was still a wild ride. It was not as scary as it could have been, but a shout out to the team that helped me resolve this, and the teams I held up from writing code, for being patient as I waded through a mess of my creation. We all came right-side up at the end, and if we don’t learn from our mistakes, then we’re not doing it right.
This article is originated from https://sethmccombs.github.io/work/2018/12/03/Outages.html
Seth McCombs is a Tigera guest blogger. He is a SRE/DevOps/Infrastructure engineer and all around container advocate, with a love of Open Source and Cloud Native.
Free Online Training
Access Live and On-Demand Kubernetes Tutorials
Calico Enterprise – Free Trial
Solve Common Kubernetes Roadblocks and Advance Your Enterprise Adoption
Get updates on blog posts, new releases and more!