When things go wrong on a large website, it can be fun to read the dramatic stories of high pressure incidents where nothing goes as planned. It makes for good reading. Every once in a while though, we get a success story. Every once in a while, things go exactly as planned.

GitLab.com is a large, high availability instance of GitLab. It is maintained by the Infrastructure group, which currently consists of 20 to 24 engineers (depending on how you count), four managers, and a director, distributed all around the world. Distributed, in this case, does not mean across a few different offices. There are three or four major cities which have more than one engineer but with the exception of coworking days nobody is working from the same building.

In order to handle the load generated by about four million users working on around 12 million projects, GitLab.com breaks out the individual components of the GitLab product and currently spreads them out over 271 production servers.

The site is slowly migrating to using Hashicorp's Consul for service location. Consul can be thought of like DNS, in that it associates a well-known name with the actual physical location of that service. It also provides other useful functions such as storing dynamic configuration for services, as well as locking for clusters. All of the Consul client and server components talk to each other over encrypted connections. These connections require a certificate at each end to validate the identity of the client and server and to provide the encryption key. The main component of GitLab.com which currently relies on this service is the database and its high availability system Patroni. Like any website that provides functionality and not just information, the database is the central service that everything else depends on. Without the database, the website, API, CI pipelines, and git services will all deny requests and return errors.


The issue came to our attention when a database engineer noticed that one of our database servers in the staging environment could not reconnect to the staging Consul server after the database node was restarted.

It turns out that the TLS certificate was expired. This is normally a simple fix. Someone would go to the Certificate Authority (CA) and request a renewal – or if that fails, generate a new certificate to be signed by the same CA. That certificate would replace the expired copy and the service would be restarted. All of the connections should reestablish using the new certificate and just like with any other rolling configuration change, it should be transparent to all users.

After looking everywhere, and asking everyone on the team, we got the definitive answer that the CA key we created a year ago for this self-signed certificate had been lost.

These test certificates were generated for the original proof-of-concept installation for this service and were never intended to be transitioned into production. However, since everything was working perfectly, the expired test certificate had not been calling attention to itself. A few things should have been done, including: Rebuilding the service with production in mind; conducting a production readiness review; and monitoring. But a year ago, our production team was in a very different place. We were small with just four engineers, and three new team members: A manager, director, and engineer, all of whom were still onboarding. We were less focused on the gaps that led to this oversight a year ago and more focused on fixing the urgent problem today.

Validating the problem

First, we needed to validate the problem using the information we'd gathered. Since we couldn't update the existing certificates, we turned validation off on the client that couldn't connect. Turning validation off didn't change anything since the encrypted connections validate both the cluster side and client side. Next, we changed the setting on one server node in the cluster and so the restarted client could then connect to the server node. The problem now was that the server could no longer connect to any other cluster node and could not rejoin the cluster. The server we changed was not validating connections, meaning it was ignoring the expired certificate of its peers in the cluster but the peers were not returning the favor. They were shunning it, putting the whole cluster in a degraded state.

We realized that no matter what we did, some servers and some clients would not be able to connect to each other until after the change had been made everywhere and after every service was restarted. Unfortunately, we were talking about 255 of our 271 servers. Our tool set is designed for gradual rollouts, not simultaneous actions.

We were unsure why the site was even still online because if the clients and services could not connect it was unclear why anything was still working. We ran a small test, confirming the site was only working because the connections were already established when the certificates expired. Any interruption of these long-running connections would cause them to revalidate the new connections, resulting in them rejecting all new connections across the fleet.

Effectively, we were in the middle of an outage that had already started, but hadn't yet gotten to the point of taking down the site.

Testing in staging

We declared an incident and began testing every angle we could think of in the staging environment, including:

Everything we tried indicated that we had to break those existing connections in order to activate any change, and that we could only avoid downtime if that happened on ALL nodes at precisely the same time.

Every problem uncovered other problems and as we were troubleshooting one of our production Consul servers became unresponsive, disconnected all SSH sessions, and would not allow anyone to reconnect. The server did not log any errors. It was still sending monitoring data and was still participating in the Consul cluster. If we restarted the server, then it would not have been able to reconnect to its peers and we would have an even number of nodes. Not having quorum in the cluster would have been dangerous when we went to restart all of the nodes, so we left it in that state for the moment.


Once the troubleshooting was finished it was time to start planning.

There were a few ways to solve the problem. We could:

All of these options would incur the same risks and involve the same risky restart of all services at once.

We picked the last option. Our reasoning was that disabling the validation would eliminate the immediate risk and give us time to slowly roll out a properly robust solution in the near future, without having to worry about disrupting the whole system. It was also the smallest and most incremental change.

Working asynchronously to tackle the problem

While there was some time pressure due to the risk of network connections being interrupted, we had to consider the reality of working across timezones as we planned our solution.

We decided not to hand it off to the European shift, who were coming online soon. Being a globally distributed team, we had already handed things off from the end of the day in Mongolia, through Eastern and Western Europe and across the Americas, and were approaching the end of the day in Hawaii and New Zealand.

Australia still had a few more hours and Mongolia had started the day again, but the folks who had been troubleshooting it throughout the day had a pretty good handle on what needed to happen and what could go wrong. It made sense for them to be the ones to do the work. We decided to make a "Break Glass" plan instead. This was a merge request with all of the changes and information necessary for the European shift to get us back into a good state in case a full outage happened before anyone who had been working on it woke up. Everyone slept better knowing that we had a plan that would work even if it could not be executed without causing down time. If we were already experiencing down time, there would be no problem.

Designing our approach

In the morning (HST) everything was how we left it so we started planning how to change the settings and restart all of the services without downtime. Our normal management tools were out because of the time it takes to roll out changes. Even sequential tools such as knife ssh, mussh, or ansible wouldn't work because the change had to be precisely simultaneous. Someone joked about setting it up in cron which led us to the standard linux at command (a relative of the more widely used batch). cron would require cleanup afterward but an at command can be pushed out ahead of time with a sequential tool and will run a command at a precise time on all machines. Back in the days of hands-on, bare metal system administration, it was a useful trick for running one-time maintenance in the middle of the night or making it look like you were working when you weren't. Now at has become more obscure with the trend toward managing fleets of servers rather than big monolithic central machines. We chose to run the command sudo systemctl restart consul.service. We tested this in staging to verify that our Ubuntu distribution made environment variables like $PATH available, and that sudo did not ask for a password. On some distributions (older CentOS especially) this is not always the case.

With those successful tests, we still needed to change the config files. Luckily, there is nothing that prevents changing these ahead of time since the changes aren't picked up until the service restarts. We didn't want to do this step at the same time as the service restart so we could validate the changes and keep the at command as small as possible. We decided not to use Chef to push out the change because we needed complete and immediate transparency. Any nodes that did not get the change would fail after the restart. mussh was the tool that offered the most control and visibility while still being able to change all hosts with one command.

We also had to disable the Chef client so that it didn't overwrite the changes between when they were written and when the service restarted.

Before running anything we also needed to address the one Consul server that we couldn't access. It likely just needed to be rebooted and would come up and be unable to reconnect to the cluster. The best option was to do this manually just before starting the rest of the procedure.

Once we had mapped out the plan we practiced it in the disaster recovery environment. We used the disaster recovery environment instead of the staging environment because all of the nodes in the staging environment had already been restarted, so there were no long-running connections to test. Making the disaster recovery environment was the next best option. It did not go perfectly since the database in this environment was already in an unhealthy state but it gave us valuable information to adjust the plan.


A moment of panic

It was almost time to fix the inaccessible Consul node. The team connected in to one of the other nodes to monitor and watch logs. Suddenly, the second node started disconnecting people. It was behaving exactly like the inaccessible node had the previous day. 😱 Suspiciously, it didn't disconnect everyone. Those who were still logged in noticed that sshguard was blocking access to some of the bastion servers that all of our ssh traffic flows through when accessing the internal nodes: Infrastructure#7484. We have three bastion servers, and two were blocked because so many of us connected so many sessions so quickly. Disabling sshguard allowed everyone back in and that information was the hint we needed to manually find the one bastion which hadn't yet been blocked. It got us back into the original problem server. Disabling sshguard there left us with a fully functional node and with the ability to accept the at command to restart the Consul service at exactly the same time as the others.

We verified that we had an accurate and instantaneous way to monitor the state of the services. Watching the output of the consul operator raft list-peers command every second gave us view that looked like this:

Node                Address          State     Voter  RaftProtocol
consul-01-inf-gprd  follower  true   3
consul-03-inf-gprd  leader    true   3
consul-05-inf-gprd  follower  true   3
consul-04-inf-gprd  follower  true   3
consul-02-inf-gprd  follower  true   3

More nodes, more problems

Even the most thorough plans always miss something. At this point we realized that one of the three pgbouncer nodes which direct traffic to the correct database instance was not showing as healthy in the load balancer. One is normally in this state as a warm spare, but one of the side effects of disconnecting the pgbouncer nodes from Consul is that they would all fail their load balancer health checks. If all health checks are failing, GCP load balancers send requests to ALL nodes as a safety feature. This would lead to too many connections to our database servers, causing unintended consequences. We worked around this by removing the unhealthy node from the load balancer pool for the remainder of this activity.

This might seem like a lot of steps but without any of these prerequisites the whole process would fail. Once all of that was done, everything was ready to go.


In the end, we scheduled a maintenance window and distilled all of the research and troubleshooting down to the steps in this issue.

Everything was staged and it was time to make the changes. This course of action included four key steps. First, we paused the Patroni database high availability subsystem. Pausing would freeze database failover and keep the high availability configuration static until we were done. It would have been bad if we had a database failure during this time so minimizing the amount of time in this state was important.

Next, we ran a script on every machine that stopped the Chef client service and then changed the verify lines in the config files from true to false. It wouldn't help to have Chef trying to reconfigure anything as we made changes. We did this using mussh in batches of 20 servers at a time. Any more in parallel and our SSH agent and Yubikeys may not have been able to keep up. We were not expecting change in the state of anything from this step. The config files on disk should have the new values but the running services wouldn't change, and more importantly, no TCP connections would disconnect. That was what we got so it was time for some verification.

Our third step was to check all of the servers and a random sampling of client nodes to make sure config files had been modified appropriately. It was also a good time to double-check that the Chef client was disabled. This check turned out to be a good thing to do, because there were a few nodes that still had the Chef client active. It turned out that those nodes were in the middle of a run when we disabled the service, and it reenabled the service for us when the run completed. Chef can be so helpful. We disabled it manually on the few machines that were affected. This delayed our maintenance window by a few minutes, so we were very glad we didn't schedule the at commands first.

Finally, we needed to remove the inactive pgbouncer node from the load balancer, so when the load balancer went into its safety mode, it would only send traffic to the two that were in a known state. You might think that removing it from the load balancer would be enough, but since it also participates in a cluster via Consul the whole service needed to be shut down along with the health check, which the load balancer uses to determine whether to send it traffic. We made a note of the full command line from the process table, shut it down, and removed it from the pool.

The anxiety builds

Now was the moment of truth. It was 02:10 UTC. We pushed the following command to every server (20 at a time, using mussh): echo 'sudo systemctl restart consul.service' | at 02:20 – it took about four minutes to complete. Then we waited. We monitored the Consul servers by running watch -n 1 consul operator raft list-peers on each of them in a separate terminal. We bit our nails. We watched the dashboards for signs of db connection errors from the frontend nodes. We all held our breath, and watched the database for signs of distress. Six minutes is a long time to think: "It's 4am in Europe, so they won't notice" and "It's dinner time on the US west coast, maybe they won't notice". Trust me, six minutes is a really long time: "Sorry APAC users for your day, which we are about to ruin by missing something".

We counted down the last few seconds and watched. In the first second, the Consul servers all shut down, severing the connections that were keeping everything working. All 255 of the clients restarted at the same time. In the next second, we watched the servers return Unexpected response code: 500, which means "connection refused" in this case. The third second… still returning "panic now" or maybe it was "connection refused"… The fourth second all nodes returned no leader found, which meant that the connection was not being refused but the cluster was not healthy. The fifth second, no change. I'm thinking, just breathe, they were probably all discovering each other. In the sixth second, still no change: Maybe they're electing a leader? Second seven was the appropriate time for worry and panic. Then, the eighth second brought good news node 04 is the leader. All other nodes healthy and communicating properly. In the ninth second, we let out a collective (and globally distributed) exhale.

A quick assessment

Now it was time to check what damage that painfully long eight seconds had done. We went through our checklist:

We expected to see a small spike in errors, or at least some identifiable change, but there was nothing but the noise floor. This was excellent news! 🎉

Then we checked whether the database was communicating with the Consul servers. It was not. Everyone quickly turned their attention to the backend database servers. If they had been running normally and the high availability tool hadn't been paused, an unplanned failover would be the minimum outage we could have hoped for. It's likely that they would have gotten into a very bad state. We started to troubleshoot why it wasn't communicating with the Consul server, but about one minute into the change, the connection came up and everything synced. Apparently it just needed a little more time than the others. We verified everything, and when everyone was satisfied we turned the high availability back on.


Now that everything in the critical path was working as expected, we released the tension from our shoulders. We re-enabled Chef and merged the MR pinning the Chef recipes to the newer version, and the MR's CI job pushed the newer version to our Chef server. After picking a few low-impact servers, we manually kicked off Chef runs after checking the md5sum of the Consul client config files. After Chef finished, there was no change to the file, and the Chef client service was running normally again. We followed the same process on the Consul servers with the same result, and manually implemented it on the database servers, just for good measure. Once those all looked good, we used mussh to kick off a Chef run on all of the servers using the same technique we used to turn them off.

Now all that was left was to straighten everything out with pgbouncer and the database load balancer and then we could fully relax. Looking at the heath checks, we noticed that the two previously healthy nodes were not returning healthy. The health checks are used to tell the load balancer which pgbouncer nodes have a Consul lock and therefore which nodes to send the traffic. A little digging showed that after retrying to connect to the Consul service a few times, they gave up. This was not ideal, so we opened an Infrastructure issue to fix it later and restarted the health checks manually. Everything showed normal so we added the inactive node back to the load balancer. The inactive node's health check told the load balancer not to select it, and since the load balancer was no longer in failsafe mode (due to the other node's health checks succeeding) the load balancer refrained from sending it traffic.


Simultaneously restarting all of the Consul components with the new configuration put everything back into its original state, other than the validation setting which we set to false, and the TCP sessions which we restarted. After this change, the Consul clients will still be using TLS encryption but will ignore the fact that our cert is now expired. This is still not an ideal state but it gives us time to get there in a thoughtful way rather than as a rushed workaround.

Every once in a while we get into a situation that all of the fancy management tools just can't fix. There is no run book for situations such as the one we encountered. The question we were asked most frequently once people got up to speed was: "Isn't there some instructional walkthrough published somewhere for this type of thing?". For replacing a certificate from the same authority, yes definitely. For replacing a certificate on machines that can have downtime, there are plenty. But for keeping traffic flowing when hundreds of nodes need to change a setting and reconnect within a few seconds of each other… that's just not something that comes up very often. Even if someone wrote up the procedure it wouldn't work in our environment with all of the peripheral moving parts that required our attention.

In these types of situations there is no shortcut around thinking things through methodically. In this case, there were no tools or technologies that could solve the problem. Even in this new world of infrastructure as code, site reliability engineering, and cloud automation, there is still room for old fashioned system administrator tricks. There is just no substitute for understanding how everything works. We can try to abstract it away to make our day-to-day responsibilities easier, but when it comes down to it there will always be times when the best tool for the job is a solid plan.

Cover image by Thomas Jensen on Unsplash

Git is a trademark of Software Freedom Conservancy and our use of 'GitLab' is under license