Logos

This post was originally published in the Wikimedia Tech blog, and is authored by Arturo Borrero Gonzalez and Brooke Storm.

One of the most successful and important products provided by the Wikimedia Cloud Services team at the Wikimedia Foundation is Toolforge. Toolforge is a platform that allows users and developers to run and use a variety of applications that help the Wikimedia movement and mission from the technical point of view in general. Toolforge is a hosting service commonly known in the industry as a Platform as a Service (PaaS). Toolforge is powered by two different backend engines, Kubernetes and GridEngine.

This article focuses on how we made a better Toolforge by integrating a newer version of Kubernetes and, along with it, some more modern workflows.

The starting point in this story is 2018. Yes, two years ago! We identified that we could do better with our Kubernetes deployment in Toolforge. We were using a very old version, v1.4. Using an old version of any software has more or less the same consequences everywhere: you lack security improvements and some modern key features.

Once it was clear that we wanted to upgrade our Kubernetes cluster, both the engineering work and the endless chain of challenges started.

It turns out that Kubernetes is a complex and modern technology, which adds some extra abstraction layers to add flexibility and some intelligence to a very old systems engineering need: hosting and running a variety of applications.

Our first challenge was to understand what our use case for a modern Kubernetes was. We were particularly interested in some key features:

  • The increased security and controls required for a public user-facing service, using RBAC, PodSecurityPolicies, quotas, etc.
  • Native multi-tenancy support, using namespaces
  • Advanced web routing, using the Ingress API

Soon enough we faced another Kubernetes native challenge: the documentation. For a newcomer, learning and understanding how to adapt Kubernetes to a given use case can be really challenging. We identified some baffling patterns in the docs. For example, different documentation pages would assume you were using different Kubernetes deployments (Minikube vs kubeadm vs a hosted service). We are running Kubernetes like you would on bare-metal (well, in CloudVPS virtual machines), and some documents directly referred to ours as a corner case.

During late 2018 and early 2019, we started brainstorming and prototyping. We wanted our cluster to be reproducible and easily rebuildable, and in the Technology Department at the Wikimedia Foundation, we rely on Puppet for that. One of the first things to decide was how to deploy and build the cluster while integrating with Puppet. This is not as simple as it seems because Kubernetes itself is a collection of reconciliation loops, just like Puppet is. So we had to decide what to put directly in Kubernetes and what to control and make visible through Puppet. We decided to stick with kubeadm as the deployment method, as it seems to be the more upstream-standardized tool for the task. We had to make some interesting decisions by trial and error, like where to run the required etcd servers, what the kubeadm init file would look like, how to proxy and load-balance the API on our bare-metal deployment, what network overlay to choose, etc. If you take a look at our public notes, you can get a glimpse of the number of decisions we had to make.

Our Kubernetes wasn’t going to be a generic cluster, we needed a Toolforge Kubernetes service. This means we don’t use some of the components, and also, we add some additional pieces and configurations to it. By the second half of 2019, we were working full-speed on the new Kubernetes cluster. We already had an idea of what we wanted and how to do it.

There were a couple of important topics for discussions, for example:

  • Ingress
  • Validating admission controllers
  • Security policies and quotas
  • PKI and user management

We will describe in detail the final state of those pieces in another blog post, but each of the topics required several hours of engineering time, research, tests, and meetings before reaching a point in which we were comfortable with moving forward.

By the end of 2019 and early 2020, we felt like all the pieces were in place, and we started thinking about how to migrate the users, the workloads, from the old cluster to the new one. This migration plan mostly materialized in a Wikitech page which contains concrete information for our users and the community.

The interaction with the community was a key success element. Thanks to our vibrant and involved users, we had several early adopters and beta testers that helped us identify early flaws in our designs. The feedback they provided was very valuable for us. Some folks helped solve technical problems, helped with the migration plan or even helped make some design decisions. Worth noting that some of the changes that were presented to our users were not easy to handle for them, like new quotas and usage limits. Introducing new workflows and deprecating old ones is always a risky operation.

Even though the migration procedure from the old cluster to the new one was fairly simple, there were some rough edges. We helped our users navigate them. A common issue was a webservice not being able to run in the new cluster due to stricter quota limiting the resources for the tool. Another example is the new Ingress layer failing to properly work with some webservices’s particular options.

By March 2020, we no longer had anything running in the old Kubernetes cluster, and the migration was completed. We then started thinking about another step towards making a better Toolforge, which is introducing the toolforge.org domain. There is plenty of information about the change to this new domain in Wikitech News.

The community wanted a better Toolforge, and so do we, and after almost 2 years of work, we have it! All the work that was done represents the commitment of the Wikimedia Foundation to support the technical community and how we really want to pursue technical engagement in general in the Wikimedia movement. In a follow-up post we will present and discuss more in-depth about some technical details of the new Kubernetes cluster, stay tuned!

This post was originally published in the Wikimedia Tech blog, and is authored by Arturo Borrero Gonzalez and Brooke Storm.