Kubernetes

A Comprehensive Overview

Kubernetes v1.8
Agenda

● Introduction ● Concepts
○ Who am I? ○ Core
○ What is Kubernetes? ○ Workloads
○ What does Kubernetes do? ○ Network
● Architecture ○ Storage
○ Master Components ○ Configuration
○ Node Components ○ Auth and Identity
○ Additional Services ● Behind the Scenes
○ Networking ○ Deployment from Beginning to End
Introduction
Intro - Who am I?

Bob Killen / rkillen@umich.edu

Twitter / Github: @mrbobbytables

Senior Research Cloud Administrator @ ARC-TS

http://arc-ts.umich.edu
Intro - What is Kubernetes?

Kubernetes or K8s was a project spun out of Google as a open source

next-gen container scheduler designed with the lessons learned from
developing and managing Borg and Omega.

Kubernetes was designed from the ground-up as a loosely coupled collection

of components centered around deploying, maintaining, and scaling
applications.
Intro - What Does Kubernetes do?

Kubernetes is the linux kernel of distributed systems.

It abstracts away the underlying hardware of the nodes and provides a

uniform interface for applications to be both deployed and consume the
shared pool of resources.
Kubernetes
Architecture
Architecture Overview

Masters - Acts as the primary control plane for Kubernetes. Masters are
responsible at a minimum for running the API Server, scheduler, and cluster
controller. They commonly also manage storing cluster state, cloud-provider
specific components and other cluster essential services.

Nodes - Are the ‘workers’ of a Kubernetes cluster. They run a minimal agent
that manages the node itself, and are tasked with executing workloads as
designated by the master.
Architecture
Overview
Master
Components
Master Components

● Kube-apiserver
● Etcd
● Kube-controller-manager
● Cloud-controller-manager
● Kube-scheduler
kube-apiserver

The apiserver provides a forward facing REST interface into the kubernetes
control plane and datastore. All clients, including nodes, users and other
applications interact with kubernetes strictly through the API Server.

It is the true core of Kubernetes acting as the gatekeeper to the cluster by

handling authentication and authorization, request validation, mutation, and
admission control in addition to being the front-end to the backing datastore.
etcd

Etcd acts as the cluster datastore; providing a strong, consistent and highly
available key-value store used for persisting cluster state.
kube-controller-manager

The controller-manager is the primary daemon that manages all core

component control loops. It monitors the cluster state via the apiserver and
steers the cluster towards the desired state.

List of core controllers:

https://github.com/kubernetes/kubernetes/blob/master/cmd/kube-controller-manager/app/controllermanager.go#L332
cloud-controller-manager

The cloud-controller-manager is a daemon that provides cloud-provider

specific knowledge and integration capability into the core control loop of
Kubernetes. The controllers include Node, Route, Service, and add an
additional controller to handle PersistentVolumeLabels .
kube-scheduler

Kube-scheduler is a verbose policy-rich engine that evaluates workload

requirements and attempts to place it on a matching resource. These
requirements can include such things as general hardware reqs, affinity,
anti-affinity, and other custom resource requirements.
Node
Components
Node Components

● Kubelet
● Kube-proxy
● Container runtime engine
kubelet

Acts as the node agent responsible for managing pod lifecycle on its host.
Kubelet understands YAML container manifests that it can read from several
sources:

● File path
● HTTP Endpoint
● Etcd watch acting on any changes
● HTTP Server mode accepting container manifests over a simple API.
kube-proxy

Manages the network rules on each node and performs connection

forwarding or load balancing for Kubernetes cluster services.

Available Proxy Modes:

● Userspace
● iptables
● ipvs (alpha in 1.8)
Container Runtime

With respect to Kubernetes, A container runtime is a CRI (Container Runtime Interface)

compatible application that executes and manages containers.

● Containerd (docker)
● Cri-o
● Rkt
● Kata (formerly clear and hyper)
● Virtlet (VM CRI compatible runtime)
Additional Services

Kube-dns - Provides cluster wide DNS Services. Services are resolvable to

<service>.<namespace>.svc.cluster.local.

Heapster - Metrics Collector for kubernetes cluster, used by some resources

such as the Horizontal Pod Autoscaler. (required for kubedashboard metrics)

Kube-dashboard - A general purpose web based UI for kubernetes.

Networking
Networking - Fundamental Rules

1) All Pods can communicate with all other Pods without NAT
2) All nodes can communicate with all Pods (and vice-versa) without NAT.
3) The IP that a Pod sees itself as is the same IP that others see it as.
Networking - Fundamentals Applied

Containers in a pod exist within the same network namespace and share an
IP; allowing for intrapod communication over localhost.

Pods are given a cluster unique IP for the duration of its lifecycle, but the pods
themselves are fundamentally ephemeral.

Services are given a persistent cluster unique IP that spans the Pods lifecycle.

External Connectivity is generally handed by an integrated cloud provider or

other external entity (load balancer)
Networking - CNI

Networking within Kubernetes is plumbed via the Container Network

Interface (CNI), an interface between a container runtime and a network
implementation plugin.

Compatible CNI Network Plugins:

● Calico ● kube-router
● Cillium ● Multus
● Contiv ● OpenVSwitch
● Contrail ● OVN
● Flannel ● Romana
● GCE ● Weave
Kubernetes
Concepts
Kubernetes Concepts - Core

Cluster - A collection of hosts that aggregate their available resources including cpu, ram, disk,
and their devices into a usable pool.

Master - The master(s) represent a collection of components that make up the control plane of
Kubernetes. These components are responsible for all cluster decisions including both
scheduling and responding to cluster events.

Node - A single host, physical or virtual capable of running pods. A node is managed by the
master(s), and at a minimum runs both kubelet and kube-proxy to be considered part of the
cluster.

Namespace - A logical cluster or environment. Primary method of dividing a cluster or

scoping access.
Concepts - Core (cont.)

Label - Key-value pairs that are used to identify, describe and group together related sets of
objects. Labels have a strict syntax and available character set. *

Annotation - Key-value pairs that contain non-identifying information or metadata.

Annotations do not have the the syntax limitations as labels and can contain structured or
unstructured data.

Selector - Selectors use labels to filter or select objects. Both equality-based (=, ==, !=) or
simple key-value matching selectors are supported.

* https://kubernetes.io/docs/concepts/overview/working-with-objects/labels/#syntax-and-character-set
 Labels, and Annotations,
and Selectors

Labels:
app: nginx
tier: frontned

Annotations
description: “nginx frontend”

Selector:
app: nginx
tier: frontend
 Set-based selectors

Valid Operators:
● In
● NotIn
● Exists
● DoesNotExist

Supported Objects with set-based selectors:

● Job
● Deployment
● ReplicaSet
● DaemonSet
● PersistentVolumeClaims
Concepts - Workloads

Pod - A pod is the smallest unit of work or management resource within Kubernetes. It is
comprised of one or more containers that share their storage, network, and context
(namespace, cgroups etc).

ReplicationController - Method of managing pod replicas and their lifecycle. Their

scheduling, scaling, and deletion.

ReplicaSet - Next Generation ReplicationController. Supports set-based selectors.

Deployment - A declarative method of managing stateless Pods and ReplicaSets. Provides

rollback functionality in addition to more granular update control mechanisms.
Deployment
ReplicaSet
Contains configuration
Generated ReplicaSet
of how updates or
from Deployment spec.
‘deployments’ should be
managed in addition to
the pod template used to
generate the ReplicaSet.
Concepts - Workloads (cont.)

StatefulSet - A controller tailored to managing Pods that must persist or maintain state. Pod
identity including hostname, network, and storage will be persisted.

DaemonSet - Ensures that all nodes matching certain criteria will run an instance of a
supplied Pod. Ideal for cluster wide services such as log forwarding, or health monitoring.
 StatefulSet

● Attaches to ‘headeless service’ (not shown) nginx.

● Pods given unique ordinal names using the pattern
<statefulset name>-<ordinal index>.
● Creates independent persistent volumes based on
the ‘volumeClaimTemplates’.
 DaemonSet

● Bypasses default scheduler

● Schedules a single instance on every host while
adhering to tolerances and taints.
Concepts - Workloads (cont.)

Job - The job controller ensures one or more pods are executed and successfully terminates. It
will do this until it satisfies the completion and/or parallelism condition.

CronJob - An extension of the Job Controller, it provides a method of executing jobs on a

cron-like schedule.
 Jobs

● Number of pod executions can be controlled

via spec.completions
● Jobs can be parallelized using spec.parallelism
● Jobs and Pods are NOT automatically
cleaned up after a job has completed.
 CronJob

● Adds cron schedule to job template

Concepts - Network

Service - Services provide a method of exposing and consuming L4 Pod network accessible
resources. They use label selectors to map groups of pods and ports to a cluster-unique virtual
IP.

Ingress - An ingress controller is the primary method of exposing a cluster service (usually
http) to the outside world. These are load balancers or routers that usually offer SSL
termination, name-based virtual hosting etc.
 Service

● Acts as the unified method of accessing replicated pods.

● Four major Service Types:
○ CluterIP - Exposes service on a strictly cluster-internal IP (default)
○ NodePort - Service is exposed on each node’s IP on a statically
defined port.
○ LoadBalancer - Works in combination with a cloud provider to
expose a service outside the cluster on a static external IP.
○ ExternalName - used to references endpoints OUTSIDE the cluster
by providing a static internally referenced DNS name.
 Ingress Controller

● Deployed as a pod to one or more hosts

● Ingress controllers are an external
controller with multiple options.
○ Nginx
○ HAproxy
○ Contour
○ Traefik
● Specific features and controller specific
configuration is passed through
annotations.
Concepts - Storage

Volume - Storage that is tied to the Pod Lifecycle, consumable by one or more
containers within the pod.

PersistentVolume - A PersistentVolume (PV) represents a storage resource. PVs are

commonly linked to a backing storage resource, NFS, GCEPersistentDisk, RBD etc. and are
provisioned ahead of time. Their lifecycle is handled independently from a pod.

PersistentVolumeClaim - A PersistentVolumeClaim (PVC) is a request for storage that

satisfies a set of requirements instead of mapping to a storage resource directly. Commonly
used with dynamically provisioned storage.

StorageClass - Storage classes are an abstraction on top of an external storage resource.

These will include a provisioner, provisioner configuration parameters as well as a PV
reclaimPolicy.
Volumes
 Persistent Volumes

● PVs are a cluster-wide resource

● Not directly consumable by a Pod
● PV Parameters:
○ Capacity
○ accessModes
■ ReadOnlyMany (ROX)
■ ReadWriteOnce (RWO)
■ ReadWriteMany (RWX)
○ persistentVolumeReclaimPolicy
■ Retain
■ Recycle
■ Delete
○ StorageClass
Persistent Volume Claims

● PVCs are scoped to namespaces

● Supports accessModes like PVs
● Uses resource request model similar to Pods
● Claims will consume storage from matching PVs
or StorageClasses based on storageClass and
selectors.
 Storage Classes

● Uses an external system defined by the

provisioner to dynamically consume and
allocate storage.
● Storage Class Fields
○ Provisioner
○ Parameters
○ reclaimPolicy
Concepts - Configuration

ConfigMap - Externalized data stored within kubernetes that can be referenced as a

commandline argument, environment variable, or injected as a file into a volume mount. Ideal
for separating containerized application from configuration.

Secret - Functionally identical to ConfigMaps, but stored encoded as base64, and encrypted at
rest (if configured).
ConfigMaps and Secrets

● Can be used in Pod Config:

○ Injected as a file
○ Passed as an environment variable
○ Used as a container command (requires passing as env var)
Concepts - Auth and Identity (RBAC)

[Cluster]Role - Roles contain rules that act as a set of permissions that apply verbs like “get”,
“list”, “watch” etc over resources that are scoped to apiGroups. Roles are scoped to namespaces,
and ClusterRoles are applied cluster-wide.

[Cluster]RoleBinding - Grant the permissions as defined in a [Cluster]Role to one or more

“subjects” which can be a user, group, or service account.

ServiceAccount- ServiceAccounts provide a consumable identity for pods or external

services that interact with the cluster directly and are scoped to namespaces.
 [Cluster]Role

● Permissions translate to url

path. With “” defaulting to core
group.

● Resources act as items the role

should be granted access to.

● Verbs are the actions the role

can perform on the referenced
resources.
[Cluster]RoleBinding

● Can reference multiple subjects

● Subjects can be of kind:
○ User
○ Group
○ ServiceAccount
● roleRef targets a single role only.
Behind
The Scenes
Behind
The Scenes
Deployment From
Beginning to End
Kubectl

1) Kubectl performs client side

validation on manifest (linting).

2) Manifest is prepared and serialized

creating a JSON payload.
APIserver Request Loop

3) Kubectl authenticates to apiserver via x509, jwt,

http auth proxy, other plugins, or http-basic auth.

4) Authorization iterates over available AuthZ

sources: Node, ABAC, RBAC, or webhook.

5) AdmissionControl checks resource quotas,

other security related checks etc.

6) Request is stored in etcd.

7) Initializers are given opportunity to mutate request before the object is published.

8) Request is published on apiserver.

Deployment Controller

9) Deployment Controller is notified of the new

Deployment via callback.

10) Deployment Controller evaluates cluster state and

reconciles the desired vs current state and forms a
request for the new ReplicaSet.

11) apiserver request loop evaluates Deployment

Controller request.

12) ReplicaSet is published.

ReplicaSet Controller

13) ReplicaSet Controller is notified of the new ReplicaSet

via callback.

14) ReplicaSet Controller evaluates cluster state and

reconciles the desired vs current state and forms a request
for the desired amount of pods.

15) apiserver request loop evaluates ReplicaSet

Controller request.

16) Pods published, and enter ‘Pending’ phase.

Scheduler

17) Scheduler monitors published pods with no

‘NodeName’ assigned.

18) Applies scheduling rules and filters to find a

suitable node to host the Pod.

19) Scheduler creates a binding of Pod to Node and

POSTs to apiserver.

20) apiserver request loop evaluates POST request.

21) Pod status is updated with node binding and sets

status to ‘PodScheduled’.
Kubelet - PodSync

22) The kubelet daemon on every node polls the apiserver filtering
for pods matching its own ‘NodeName’; checking its current state
with the desired state published through the apiserver.

23) Kubelet will then move through a series of internal processes to

prepare the pod environment. This includes pulling secrets,
provisioning storage, applying AppArmor profiles and other various
scaffolding. During this period, it will asynchronously be POST’ing
the ‘PodStatus’ to the apiserver through the standard apiserver
request loop.
Pause and Plumbing

24) Kubelet then provisions a ‘pause’ container via the

CRI (Container Runtime Interface). The pause container
acts as the parent container for the Pod.

25) The network is plumbed to the Pod via the CNI

(Container Network Interface), creating a veth pair
attached to the pause container and to a container
bridge (cbr0).

26) IPAM handled by the CNI plugin assigns an IP to the

pause container.
Kublet - Create Containers

24) Kubelet pulls the container Images.

25) Kubelet first creates and starts any init containers.

26) Once the optional init containers complete, the

primary pod containers are started.
Pod Status

27) If there are any liveless/readiness probes, these are executed before the
PodStatus is updated.

28) If all complete successfully, PodStatus is set to ready and the container
has started successfully.

The Pod is Deployed!

Questions?