As part of SCaLE 15x, I took part in the first Open Source Infra Day where a number of other sysadmins and I shared stories and patterns which have helped us maintain open source infrastructure. As part of the "unconference" tracks, I suggested and then led the session "Running containers in production." As my luck would have it, in a group of roughly 10 people representing various groups, Jenkins was the only project running production services in containers. I thought I should share what it's like, and why you should stop standing on the sidelines and give containers in production a try.
Containers have suffered from an unreasonable amount of hype. Which has caused many people, myself included, to be exceptionally skeptical of their maturity and utility in a modern infrastructure. Here's what containers, by themselves, do not solve:
However, here is what containers can and do solve for:
At this point, I won't not use containers anymore. Their benefits far outweigh their flaws, even in production environments. But there are flaws.
The advantages of using containers are fairly well-documented across various presentations, blog posts, and gushy tweets. For my usage, which is partly personal and partly for the Jenkins project, the benefits are as follows.
For a mostly mixed infrastructure like that of the Jenkins project, this can be hugely beneficial. For some applications newer, bleeding edge Java Runtime Environments have been required, whilst others might need whatever basic JRE is available. With the container including that form of system dependency within it, the virtual machines running the containers don't need to know, or care, about the mixed application runtime requirements and dependencies.
Despite hype to the contrary, Operations is still a necessary specialization in most organizations, the same as design, quality engineering, or product management. That doesn't necessitate Ops "ownership" of applications, or their implementations. The more freedom, and responsibility, which can be granted to application developers the faster they will be able to build, test, and deploy their applications.
In the case of an open source project like Jenkins, this is especially true. Very few contributors have the experience, and trust, necessary to act as infrastructure administrators. Those contributors do not have the time to tend to, or "own", each application.
This defined hand-off point gave the application developer a tangible achievable goal, which is within his reach, to meet in order for his application to be deployed.
As much as I would like to teach more people how to use Puppet, or Chef, most developers have enough to cram into their brains without adding configuration management to the mix.
By endeavouring to build the application into a container, the application developer can, for the most part, run the container locally just as it would run in the production environment. Not only does this make local development easier, but it also empowers developers to take responsibility for and modify the runtime environment to a greater degree than if the application ran differently locally versus in production.
NOTE: By constructing hellish webs of inter-container dependencies, or containers which require excessive amounts of environment variables, etc, this advantage will be lost. A poorly designed or implemented application is still going to be difficult to run, regardless of whether it's packaged in a container or not.
Continuous delivery of containers is most certainly a topic for a blog post unto itself, but suffice it to say, it's amazingly easy with Jenkins these days. The key benefit containers provide to a continuous delivery process is immutability. Building a container provides an immutable object which can progress through a pipeline towards production.
While this is also feasible with a
.war file, or other application archive,
the inclusion of the runtime environment within the package ensures that the
runtime characteristics between a "testing", "staging", and "production"
environments are all the same.
Jenkinsfile for the application referenced
above, one stage of the
defined Jenkins Pipeline builds the
container, then the next stage runs the container and performs some
rudimentary acceptance tests against that container. On at least one occasion
this has prevented the deployment of a genuinely broken application.
The momentum behind the container ecosystem continues to grow, which means a beleagured infrastructure administrator, like yours truly, can take advantage of a myriad of tools and technologies to make life easier. Technologies like Kubernetes, especially on Azure or Google Container Engine (GKE), abstract a lot away, in a good way, to where you can think of the service as opposed to the specific implementation details of the application
For the Jenkins project's next major iteration of infrastructure on Azure, we're deploying a Kubernetes environment to deploy all applications into, thereby dramatically reducing the runtime footprint, and cost, of the dozen little apps floating around. As part of that migration, we're deploying Fluentd for log aggregation, and Datadog's Kubernetes support for thorough monitoring of applications as well. All of this tooling comes together to allow us to describe services, rather than processes, which are necessary for the business of the project.
I believe, and now have proof to back it up, that by supporting containers in production we (the Jenkins project) are able to support more varied applications, with faster and more automated release cycles, than without containers. If speed and reliability are not important to you, then it's probably safe to continue not running containers in production.
There is no such thing as free lunch, and as much as I would like to say everything in containers is sunshine and rainbows, there are definitely some growing pains, and issues with containers in production.
I will generalize here, because there is no "one container networking" mechanism. Host networks, overlay networks, weave networks, etc. There's a lot of ways to network containers together, whether in a cluster like Kubernetes or on a single machine. All of them rely on what I will broadly categorize as "awful kernel networking tricks."
In the pre-Kubernetes days for the Jenkins project we are still orchestrating
the deployment of containers with Puppet on individual virtual machines. This
means we're taking advantage of overlay networks and trusting that the Docker
daemon will configure the appropriate
iptables rules to expose our containers
to the external network interfaces. To put a finer point on it, this means the
Docker container is creating Network Address Translation rules (aka
Masquerading in iptables parlance) on the machine to get traffic in and out.
NAT is bad, and it should feel bad.
What we have observed over time, is that the Docker daemon cannot be trusted to
iptables rules as the container lifecycle changes. What
intermittently happens is: a service goes offline after a new deployment, because
the kernel is still NATing traffic from the external port 8080 to port 8080 on
an old virtual IP address (
22.214.171.124) instead of the new virtual IP
address the new container was deployed with (
126.96.36.199). When we look at the
iptables -t nat -L output, we see the
POSTROUTING chains with
rules for both IPs, but the older one remains with a higher precedence.
Unfortunately this is "fucking impossible" (my words) to reproduce, so I haven't been able to file a suitable bug report upstream.
This issue exposed a novel gap in our monitoring practices as well. Previously,
some applications had process checks, basically "is apache2 running? great,
everything is fine." When this issue manifests itself, the application
container is running properly, the process table will be correct, the issue
will hide in the
nat table and users will complain that the
application is unreachable.
The obvious take-away here is a known monitoring best practice: monitor how the user sees the application, not how your infrastructure sees it.
Instead of checking that the process serving the web application is running,
curl(1) against the domain name, or external IP, actually validates
that the application is online and reachable, meeting the service contract for
Running the Docker daemon,
dockerd, for long periods of time is apparently
not a common practice, but we do it. As a result we have observed, on daemons
with heavy churn or high workloads, that
dockerd periodically will
wedge/freeze/deadlock taking every container running on that daemon with it.
strace(1) shows the daemon waiting on a lock (
futex(2)) which will
Unfortunately this also is "fucking impossible" (my words) to reproduce, so I haven't been able to file a suitable bug report upstream.
This issue exposed another monitoring gap, checking that
dockerd is running
is insufficient. Our docker monitoring must execute
docker commands to see if
the daemon is responsive and doing what it should. While we haven't gone so far
as to automatically restart
dockerd when this happens, it would be relatively
straight-forward to implement.
Containers typically will have some ephemeral and mapped storage, backed by a
storage backend configured in the daemon. I have tried both
aufs and seen similar behaviors of ever increasing disk usage, regardless of
what the application is doing. The behavior I have observed, but not dug too
deeply into, is that disk space allocated for the
aufs backend, for example,
will only ever grow. It will not shrink, regardless of whether the application
is actually using the space or not. This seems to "reset" when the container is
stopped and removed however.
I believe we notice this acutely because we have long-running containers which live on long-running virtual machines.
I have seen a lot of skepticism from my peers about tools like Kubernetes and the hype surrounding them. Some of this skepticism is fair, but let me make this much clear: if you do not use a tool like Kubernetes, you will end up building a shittier version of it.
Containers being lightweight allows for over-provisioning virtual machines with multiple containers per instance, and of course it would be silly to run just one instance of an application so you need replicas, and of course you will want some form of persistent storage for your applications, and all of a sudden you're staring at the domain Kubernetes aims to address.
Additionally, managing secrets can be challenging with containers. It would be bone-headed to bake a container with production SSL certificates, or API keys, within it. Instead you will need a mechanism for injecting environment-specific (testing, staging, production) credentials into containers. Without a tool like Kubernetes, you will most certainly hack something up yourself to get secrets into your containers, and it will most certainly be ugly.
You should no longer be afraid of using contianers in production. There are certainly caveats and challenges to address, but I assert those downsides are far outweighed by the benefits to the organization. Astute readers might be wondering at this point, when I will talk about containers and security? Frankly, I don't see anything novel about security with containers. Firstly, do not assume containers provide you any additional layer of security. Secondly, the patch lifecycle, while differently managed compared to traditional package management, is still more or less the same.
I recommend investigating Kubernetes on Azure or Google Container Engine, don't waste your time with Amazon's Elastic Container Service, and start by deploying small stateless applications and slowly work your way up to beefier stateful monoliths.
But do remember, there are no silver bullets. There will be flaws and challenges, but adopting containers in production will encourage more automation and continuous delivery of applications, lead to broader delegation of responsibilities, and allow infrastructure people to focus on infrastructure rather than applications.