This is an early release. Some features are not yet fully implemented.
A distributed declarative orchestrator for services that speaks TOSCA.
Khutulun is a straightforward, flexible alternative to Kubernetes, Nomad, etc.
Its primary design goal is that the outcome of orchestration would be no different from what a sysadmin would do themselves. If you want to simply install and run a bare process on a machine, Khutulun will do that for you. If you want straightforward networking based on reserved TCP ports, Khutulun won’t do anything more than keep track of those ports for you. More complex deployments using containers, virtual machines, and virtual networks are also supported, but Khutulun’s aim is to manage complexity without getting in the way of simplicity.
The guiding paradigm is policy-driven service composition based on graph representations of the service’s topology. Khutulun provides custom discovery that allows components to find each other, form a service mesh, and modify the topologies to which they belong (“Day 2” and the “operator pattern”). There are tools for injecting discovered data into configuration files and environment variables to allow “legacy” applications to participate in the mesh.
Khutulun is modular and extensible. Resource types are handled by a cooperative ecosystem of delegates, the main delegate types being for running compute workloads, for networking, and for storage. Delegates can call other delegates. They can be implemented as in-process plugins, services, and even packaged into workloads on the cluster.
Some included resource types and delegates:
- Bare processes: self-contained or otherwise installable executables and scripts
- Containers using Podman or systemd-nspawn
- Pristine containers using Distrobox (on top of Podman) or systemd-nspawn
- Pods of containers using Podman
- Virtual machines using libvirt
- TCP port reservation with support for exposure through Firewalld
- Local or networked directory storage
Plugins can optionally wrap resources in usermode systemd units. This provides a unified admin experience as well as resilience in the case of failures and restarts.
Note that Khutulun does not demand that every container (pod) have its own IP address in an internal network. (Very much unlike Kubernetes.) If desired this feature could be implemented by a networking delegate.
Cluster formation is emergent and based on the SWIM gossip protocol, with optional support for UDP multicast for automatic mutual discovery. At the minimum you need just one “seed” host to bootstrap a cluster, but because all hosts are “masters” the cluster can survive with as little as one arbitrary host.
Khutulun doesn’t distribute its management state among hosts. Instead it simply requires that all hosts have access to the same shared filesystem. A simple NFS share should be enough even for large clusters. Coordination is handled via flock.
What about setting up the cluster hosts? Bare metal tasks like partitioning drives, installing operating systems, and configuring networking and other essential services? Or cloud tasks like provisioning virtual machines, virtual storage, and virtual networks? Simply put, that work is out of the scope of Khutulun. Use a dedicated infrastructure manager instead. Khutulun can interact with such tools, for example to allow workloads to modify their own cluster, or to use a Khutulun cluster as a dedicated “management cluster” that, well, manages the hardware of all other clusters.
Included is are plugins for Terraform and Ansible that make it easy to install Khutulun on your infrastructure while preparing it.
By the way, hosts do not have to be dedicated to Khutulun and its workloads. You can use Khutulun to manage services across many machines without having to conceptualize them as a “cloud”.
What’s wrong with Kubernetes?
Kubernetes is delightfully minimalistic, as orchestrators go, but still makes some potentially costly decisions:
- The requirement that every pod have its own IP address demands complex container networking solutions and is a substantial obstacle for inter-cluster connectivity. For some use cases the cost of this requirement is too high to bear. (Multus enables “side-loading” networking, but we still need a primary IP address on the Kubernetes control plane.) Distributed storage solutions must also be made to participate in this networking scheme.
- Kubernetes is focused on one kind of activity: pods of Docker-style containers. This means that we also require a Docker-style container image repository (either external or internal to the cluster). That’s a not-insignficant cost. And what if we don’t need or want to use containers? Sometimes we want just bare processes, or virtual machines, or even other container technologies (e.g. systemd-nspawn). (Yes, KubeVirt enables VMs on Kubernetes, but they have to awkwardly dress up as pods and participate in container networking.)
- Kubernetes’s resource data model, often represented as YAML manifests, has no relational capabilities (except ownership). But cloud workloads are all about relationships, e.g. service meshes. This lack of topological expressiveness is an obstacle to application and service modeling. (And, no, Helm charts are not graphs. Though, as an alternative, check out Turandot, which brings TOSCA to Kubernetes.) Also, the Kubernetes data model isn’t easily extensible: custom resource definitions require admin access to set up and custom resources do not behave exactly like built-in types.
- Relying on etcd for cluster state limits its usability. Etcd documents have a maximum size of ~1KB. If you need to share anything more substantial then you’ll have to deploy your own solution.
If none of the above is a problem for your use case then by all means stick to Kubernetes. Otherwise, try Khutulun, because it deliberately attempts to avoid these limitations.
TOSCA is an open standard with broad industry support. It is, as of version 2.0, a pure object-oriented language that relies on “profiles”, or type libraries, that in Khutulun can work with delegates to provide specific implementations. Khutulun comes with its own TOSCA profile and ecosystem of delegates. You are encouraged to add your own.
One of the hallmarks of TOSCA is that every service is a topological graph. Moreover, the edges of the graph are first-class citizens. This killer feature supercharges your modeling power for the cloud.
The developers of Khutulun are involved in the TOSCA community and committed to improving the standard.
Why support bare processes? Don’t containers provide better isolation?
Yes, containers indeed provide better isolation and Khutulun supports them out the box via Podman, Distrobox (on top of Podman), and systemd-nspawn.
But don’t just jump on the bandwagon, ask yourself: Is isolation really what you need for your use case? And do you understand and are willing to pay for what it costs? We are in the midst of an architectural shift towards service composition and away from component isolation. Isolation is often beneficial, and in some specific use cases even necessary, but if isolation technologies get in the way of collaboration technologies then you’re are shooting yourself in the foot. Most container technologies require you to build ready-to-run container images and stand up container image registries to store them, adding significant complexity to your development and deployment workflows. Also complex is managing container networking across clusters. If you entirely own your cluster and workloads then it might save you mountains of pain to simply use bare processes with bare networking.
Consider Distrobox as a Goldilocks solution: it provides pristine containers that provide only a minimal operating system but no workloads, so you can run your workloads there instead of on the bare host. Khutulun will handle the heavy lifting for you. The result may give you the best of both worlds.
Why not use a distributed key-value store like etcd for management state?
What’s wrong with just having a filesystem shared among all hosts? Seriously, why make things more complicated than they have to be?
Note that etcd has strict limits on the size of documents, which is an obstacle for sharing large, useful binary artifacts. That means that if you need to share large, useful binary artifacts you will need to deploy yet another storage system. Are we winning yet?
Why is there no custom Khutulun cluster installer? Why recommend using Terraform and other tools instead?
Infrastructure management is a solved problem. Let’s please not reinvent the wheel just for Khutulun to have its own opinion.
Why is it called “Khutulun”?
Khutulun (Mongolian: Хотулун) was a fabled Mongolian warrior, daughter of Kublai Khan’s cousin, Kaidu.
She was likely the inspiration for Turandot, the protagonist of Count Carlo Gozzi’s commedia dell’arte play, which in turn inspired Giacomo Puccini’s opera of the same name.
And Puccini is the TOSCA processor that drives Khutulun.
How do I pronounce “Khutulun”?
- International level: “KOO-too-loon”
- Cosmopolitan level: “CHOO-too-loon” (“ch” like in “Johann Sebastian Bach”)
- Expert level: Modern Mongolian “Хотулун” (video)