Kubernetes is Not a Hosting Platform

There’s a common theme in software engineering communities of software that’s too complex. Slack and other Electron apps are frequent targets – why do we need yet another “web browser” using 2GB of RAM when IRC worked perfectly well?

While I can empathise with the performance issues, the question often betrays a misunderstanding of the problem being solved or the target audience of the software. Slack is not designed primarily for software engineers who grew up on the internet in the 90s, it’s designed for non-engineers. People who are used to spending half their day in their email, or who want to send files to each other without having to ask corporate IT to allow larger email attachments or bump their quota on the shared drive. Slack solves these problems very well.

There’s another example of this: Kubernetes (“K8s”). Some of the common criticisms include…

  • There are too many moving parts.
  • It takes a ton of configuration to be production ready.
  • Developers need to write lots of YAML boilerplate for “simple things”.
  • systemd can do all of this.
  • The same functionality can be composed together with existing open-source tools.
  • A new startup will spend all their time figuring out K8s instead of shipping their application.

I believe this is the same phenomenon as with Slack. I think these are valid criticisms for engineers who do not need K8s, but who actually need either a traditional Linux-based application deployment, or who need a hands-off hosting platform. But these criticisms miss an understanding of the problems K8s is aiming to solve.

Kubernetes as a hosting platform?

Despite what web development trends may imply, Kubernetes isn’t about running web apps. They are a thing you can build out of the parts it provides, but at its core K8s is a bunch of state machines and dependency resolvers. If what is needed is a hosting solution for a web app, K8s will indeed be far too complex and will bring a lot of maintenance overhead with it.

In fact as a “hosting platform”, K8s asks far more questions than it answers:

  • How are apps defined? K8s works at a lower level than an “app”.
  • How are secrets used? K8s “secrets” are only obfuscated configuration, more of a solution is needed for secure production apps.
  • Where does my database live? Persistent data requires more work in K8s.
  • Monitoring? Metrics and logging? Certificate provisioning?

There are answers to all of these1, but each has an operational overhead and introduces complexity. A good hosting platform should answer some or all of these.

Those who need a hosting platform are likely to get further, faster, with either a managed hosting platform such as Heroku that will address these needs, or a simpler and more understandable system built out of well known open source components2 that allow a team to have more control and understanding at a lower level.

Kubernetes as a workload orchestrator!

If a system is too complex, if it has too many moving parts, and a more static hosting solution is failing to capture dependencies and meet requirements, this is the point at which K8s becomes useful, taking responsibility for orchestrating the many moving parts and reducing the burden on the engineering team.

In K8s the user provides the desired state and the orchestrator will progressively change things under its control until the world matches that state, and then attempt to maintain it should anything outside of its control change in the future.

There are two interesting details to highlight from this:

  • The obvious one, that if a server fails the services running on it will be moved elsewhere, giving services a level of resiliency in the face of failure.
  • The less obvious one, that an engineer making changes to a system doesn’t have to usefully reason about the impact of their changes – K8s will do this for them – they can let the constraint resolution do its work and check the output.

If you can hold the state of your system in your head, if your scaling concerns come down to a single number of web server processes or queue workers, then a managed hosting platform or a more static bare-server based deployment with well known tools is likely to be a better fit.

Not only are many services in this category, but we should also be striving to create services and deployments that are simple enough that they can be deployed with simple tools.

However, when a service, or more likely a multi-service deployment, gets too complex the need for K8s arises. There are a few examples that I’ve seen that I believe illustrate where it can make a real difference. While all of these would be possible without K8s, they are examples of where it can reduce complexity3 rather than increase it.

Hey! email service

One of the main concerns for Hey! is costs. Basecamp, the company behind it, are used to low infrastructure costs as they typically host things on bare-metal servers that they manage themselves. Hey! runs in the cloud so that it can scale, but the team use K8s to manage their costs in two ways.

The first is that their services run mostly on AWS spot instances – servers that can be turned off at short notice, but which are substantially cheaper as a result. They use K8s to ensure that their service components are scaled correctly even when machines are coming and going underneath them, without interaction from engineers.

The second is by using K8s to efficiently pack services onto machines. Because services define their required resources upfront, K8s can bin-pack these services onto the available hardware. When paired with an autoscaler for the underlying server pool, this will often result in a more efficient use of resources than on a static bare-metal deployment.

OpenAI research cluster

OpenAI uses GPUs on their servers to accelerate machine learning, but GPUs and their driver setup can be flaky. They use K8s to manage the lifecycle of servers, bringing new hardware online in a testing state, running tests to check that the hardware and drivers are configured correctly, before releasing the hardware to the pool for training use.

They also use K8s primitives such as taints to implement a lightweight quota system, scheduling team work onto separate regions of the cluster, while also allowing low priority workloads to run on unused capacity from other teams. While K8s doesn’t have a particularly advanced quota system, the fact that these simple requirements could be encoded in it speaks to the flexibility of it for defining complex workflows. K8s also provides APIs and customisation points for things like more advanced quota systems to be plugged in should they be needed.

Thread’s recommendations service

Thread’s recommendation service needs up to date data about our products, in particular their stock levels. This data must be distributed to each service instance several times an hour. As well as this, we also need to ensure that there’s a minimum availability of the recommendation service based on the current load from customers and batch processing jobs.

Originally we distributed the data by pushing it to cloud storage and having a systemd timer on every server downloading the updated data on a regular schedule. This was quick to implement and easy to understand, but unfortunately failed to solve the problem. When the timer ran, all the servers would go offline at once resulting in downtime. Even after we added some random variance to the timers, we were trading off between the data being too old and being over-provisioned so that even during a dip in available servers we’d still have enough capacity.

By versioning the data as a container in our recommendation service’s pods, we’re able to treat pushing new data out to the cluster as a service deployment. This way we benefit from the K8s deployment primitives, allowing us to maintain the service at the right scale (accounting for pod autoscaling) not taking too many instances offline at the same time. K8s will also verify that instances return to service successfully and will halt a roll-out should they fail health checks.

Given that our recommendation service scales from single digits of pods to hundreds running together, across reliable and unreliable4 nodes, with code deployments potentially happening at the same time, we can push a lot of the complexity into K8s to be orchestrated for us.


When used to solve the problems it sets out to solve, Kubernetes can be a powerful component of a mature cloud service deployment. It can be used to efficiently combine requirements across scalability and reliability, and can encode workflows in a way that scales as systems become more complex.

As software engineers we are uniquely positioned to criticise software but it’s important to remember that we may not always be the target audience, and where the design choices in a system may not align with what we need, there may be those for whom it does.


  1. At Thread we use kustomize, kapp, kbld, sops, Cloud SQL, Datadog, cert-manager, and more. Each is good, but in aggregate it was a lot of work to set up.

  2. A fairly typical setup might be Ubuntu LTS, with systemd to manage services, and a tool such as Ansible to provision servers. This sort of setup is likely to be stable for years at a time, and while building such a system isn’t easy, information and guidance on this sort of server administration is plentiful.

  3. It doesn’t really reduce complexity, it just offloads it to Kubernetes’ internals. However we have APIs to hide complexity like this. Boundaries can reduce accidental complexity on all sides, and can provide a nice interface for testing. Ultimately Kubernetes is likely to have a better implementation of rolling deployments, for example, than most home-grown implementations given its extensive review, testing, and well defined semantics.

  4. We have a node pool of pre-emptible instances for batch operations in our Kubernetes cluster on Google Cloud, along side our regular node pool that runs on regular instances.