We have observed two different cluster management architectures, which can be categorized as “Borg-style” and “Mesos/Omega-style.” In the remainder of this document, we will abbreviate the latter as “Mesos-style.” Although out-of-the box Kubernetes uses a Borg-style architecture, it can also be configured in a Mesos-style architecture, and in fact can support both styles at the same time. This document describes the two approaches and describes how to deploy a Mesos-style architecture on Kubernetes.
As an aside, the converse is also true: one can deploy a Borg/Kubernetes-style architecture on Mesos.
This document is NOT intended to provide a comprehensive comparison of Borg and Mesos. For example, we omit discussion of the tradeoffs between scheduling with full knowledge of cluster state vs. scheduling using the “offer” model. That issue is discussed in some detail in the Omega paper. (See references below.)
A Borg-style architecture is characterized by:
a single logical API endpoint for clients, where some amount of processing is done on requests, such as admission control and applying defaults
generic (non-application-specific) collection abstractions described declaratively,
generic controllers/state machines that manage the lifecycle of the collection abstractions and the containers spawned from them
a generic scheduler
For example, Borg’s primary collection abstraction is a Job, and every application that runs on Borg–whether it’s a user-facing service like the GMail front-end, a batch job like a MapReduce, or an infrastructure service like GFS–must represent itself as a Job. Borg has corresponding state machine logic for managing Jobs and their instances, and a scheduler that’s responsible for assigning the instances to machines.
The flow of a request in Borg is:
Client submits a collection object to the Borgmaster API endpoint
Admission control, quota, applying defaults, etc. run on the collection
If the collection is admitted, it is persisted, and the collection state machine creates the underlying instances
The scheduler assigns a hostname to the instance, and tells the Borglet to start the instance’s container(s)
Borglet starts the container(s)
The instance state machine manages the instances and the collection state machine manages the collection during their lifetimes
Out-of-the-box Kubernetes has workload-specific abstractions (ReplicaSet, Job, DaemonSet, etc.) and corresponding controllers, and in the future may have workload-specific schedulers, e.g. different schedulers for long-running services vs. short-running batch. But these abstractions, controllers, and schedulers are not application-specific.
The usual request flow in Kubernetes is very similar, namely
Client submits a collection object (e.g. ReplicaSet, Job, …) to the API server
Admission control, quota, applying defaults, etc. run on the collection
If the collection is admitted, it is persisted, and the corresponding collection controller creates the underlying pods
Admission control, quota, applying defaults, etc. runs on each pod; if there are multiple schedulers, one of the admission controllers will write the scheduler name as an annotation based on a policy
If a pod is admitted, it is persisted
The appropriate scheduler assigns a nodeName to the instance, which triggers the Kubelet to start the pod’s container(s)
Kubelet starts the container(s)
The controller corresponding to the collection manages the pod and the collection during their lifetime
In the Borg model, application-level scheduling and cluster-level scheduling are handled by separate components. For example, a MapReduce master might request Borg to create a job with a certain number of instances with a particular resource shape, where each instance corresponds to a MapReduce worker; the MapReduce master would then schedule individual units of work onto those workers.
Mesos is fundamentally designed to support multiple application-specific “frameworks.” A framework is composed of a “framework scheduler” and a “framework executor.” We will abbreviate “framework scheduler” as “framework” since “scheduler” means something very different in Kubernetes (something that just assigns pods to nodes).
Unlike Borg and Kubernetes, where there is a single logical endpoint that receives all API requests (the Borgmaster and API server, respectively), in Mesos every framework is a separate API endpoint. Mesos does not have any standard set of collection abstractions, controllers/state machines, or schedulers; the logic for all of these things is contained in each application-specific framework individually. (Note that the notion of application-specific does sometimes blur into the realm of workload-specific, for example Chronos is a generic framework for batch jobs. However, regardless of what set of Mesos frameworks you are using, the key properties remain: each framework is its own API endpoint with its own client-facing and internal abstractions, state machines, and scheduler).
A Mesos framework can integrate application-level scheduling and cluster-level scheduling into a single component.
Note: Although Mesos frameworks expose their own API endpoints to clients, they consume a common infrastructure via a common API endpoint for controlling tasks (launching, detecting failure, etc.) and learning about available cluster resources. More details here.
Implementing the Mesos model on Kubernetes boils down to enabling application-specific collection abstractions, controllers/state machines, and scheduling. There are just three steps:
Use API plugins to create API resources for your new application-specific collection abstraction(s)
Implement controllers for the new abstractions (and for managing the lifecycle of the pods the controllers generate)
Implement a scheduler with the application-specific scheduling logic
Note that the last two can be combined: a Kubernetes controller can do the scheduling for the pods it creates, by writing node name to the pods when it creates them.
Once you’ve done this, you end up with an architecture that is extremely similar to the Mesos-style–the Kubernetes controller is effectively a Mesos framework. The remaining differences are:
In Kubernetes, all API operations go through a single logical endpoint, the API server (we say logical because the API server can be replicated). In contrast, in Mesos, API operations go to a particular framework. However, the Kubernetes API plugin model makes this difference fairly small.
In Kubernetes, application-specific admission control, quota, defaulting, etc. rules can be implemented in the API server rather than in the controller. Of course you can choose to make these operations be no-ops for your application-specific collection abstractions, and handle them in your controller.
On the node level, Mesos allows application-specific executors, whereas Kubernetes only has executors for Docker and rkt containers.
The end-to-end flow is:
Client submits an application-specific collection object to the API server
The API server plugin for that collection object forwards the request to the API server that handles that collection type
Admission control, quota, applying defaults, etc. runs on the collection object
If the collection is admitted, it is persisted
The collection controller sees the collection object and in response creates the underlying pods and chooses which nodes they will run on by setting node name
Kubelet sees the pods with node name set and starts the container(s)
The collection controller manages the pods and the collection during their lifetimes
Note: if the controller and scheduler are separated, then step 5 breaks down into multiple steps:
(5a) collection controller creates pods with empty node name.
(5b) API server admission control, quota, defaulting, etc. runs on the
pods; one of the admission controller steps writes the scheduler name as an
annotation on each pods (see pull request
#18262 for more details).
(5c) The corresponding application-specific scheduler chooses a node and writes node name, which triggers the Kubelet to start the pod’s container(s).
As a final note, the Kubernetes model allows multiple levels of iterative
refinement of runtime abstractions, as long as the lowest level is the pod. For
example, clients of application Foo might create a
FooSet which is picked up
by the FooController which in turn creates
objects, which are picked up by the BatchFoo controller and ServiceFoo
controller respectively, which in turn create pods. In between each of these
steps there is an opportunity for object-specific admission control, quota, and
defaulting to run in the API server, though these can instead be handled by the