4. Component Level Architecture

4.1. Introduction

This chapter describes in detail the Kubernetes Reference Architecture in terms of the functional capabilities and how they relate to the Reference Model requirements, i.e. how the infrastructure profiles are determined, documented and delivered.

The specifications defined in this chapter will be detailed with unique identifiers, which will follow the pattern: ra2.<section>.<index>, e.g. ra2.ch.001 for the first requirement in the Kubernetes Node section. These specifications will then be used as requirements input for the Kubernetes Reference Implementation and any vendor or community implementations.

Figure 4-1 below shows the architectural components that are described in the subsequent sections of this chapter.

Figure 4-1: Kubernetes Reference Architecture

4.2. Kubernetes Node

This section describes the configuration that will be applied to the physical or virtual machine and an installed Operating System. In order for a Kubernetes Node to be conformant with the Reference Architecture it must be implemented as per the following specifications:

Ref	Specification	Details	Requirement Trace	Reference Implementation Trace
ra2.ch.001	Huge pages	When hosting workloads matching the Network Intensive profile, it must be possible to enable Huge pages (2048KiB and 1048576KiB) within the Kubernetes Node OS, exposing schedulable resources hugepages-2Mi and hugepages-1Gi.	infra.com.cfg.004	4.3.1
ra2.ch.002	SR-IOV capable NICs	When hosting workloads matching the Network Intensive profile, the physical machines on which the Kubernetes Nodes run must be equipped with NICs that are SR-IOV capable.	e.cap.013	3.3
ra2.ch.003	SR-IOV Virtual Functions	When hosting workloads matching the Network Intensive profile, SR-IOV virtual functions (VFs) must be configured within the Kubernetes Node OS, as the SR-IOV Device Plugin does not manage the creation of these VFs.	e.cap.013	4.3.1
ra2.ch.004	CPU Simultaneous Multi-Threading (SMT)	SMT must be enabled in the BIOS on the physical machine on which the Kubernetes Node runs.	infra.hw.cpu.cfg.004	3.3
ra2.ch.005	CPU Allocation Ratio - VMs	For Kubernetes nodes running as Virtual Machines, the CPU allocation ratio between vCPU and physical CPU core must be 1:1.
ra2.ch.006	CPU Allocation Ratio - Pods	To ensure the CPU allocation ratio between vCPU and physical CPU core is 1:1, the sum of CPU requests and limits by containers in Pod specifications must remain less than the allocatable quantity of CPU resources (i.e. requests.cpu < allocatable.cpu and limits.cpu < allocatable.cpu).	infra.com.cfg.001	3.3
ra2.ch.007	IPv6DualStack	To support IPv4/IPv6 dual stack networking, the Kubernetes Node OS must support and be allocated routable IPv4 and IPv6 addresses.
ra2.ch.008	Physical CPU Quantity	The physical machines on which the Kubernetes Nodes run must be equipped with at least 2 physical sockets, each with at least 20 CPU cores.	infra.hw.cpu.cfg.001, infra.hw.cpu.cfg.002	3.3
ra2.ch.009	Physical Storage	The physical machines on which the Kubernetes Nodes run should be equipped with Sold State Drives (SSDs).	infra.hw.stg.ssd.cfg.002	3.3
ra2.ch.010	Local Filesystem Storage Quantity	The Kubernetes Nodes must be equipped with local filesystem capacity of at least 320GB for unpacking and executing containers. Note, extra should be provisioned to cater for any overhead required by the Operating System and any required OS processes such as the container runtime, Kubernetes agents, etc.	e.cap.003	3.3
ra2.ch.011	Virtual Node CPU Quantity	If using VMs, the Kubernetes Nodes must be equipped with at least 16 vCPUs. Note, extra should be provisioned to cater for any overhead required by the Operating System and any required OS processes such as the container runtime, Kubernetes agents, etc.	e.cap.001
ra2.ch.012	Kubernetes Node RAM Quantity	The Kubernetes Nodes must be equipped with at least 32GB of RAM. Note, extra should be provisioned to cater for any overhead required by the Operating System and any required OS processes such as the container runtime, Kubernetes agents, etc.	e.cap.002	3.3
ra2.ch.013	Physical NIC Quantity	The physical machines on which the Kubernetes Nodes run must be equipped with at least four (4) Network Interface Card (NIC) ports.	infra.hw.nic.cfg.001	3.3
ra2.ch.014	Physical NIC Speed - Basic Profile	The speed of NIC ports housed in the physical machines on which the Kubernetes Nodes run for workloads matching the Basic Profile must be at least 10Gbps.	infra.hw.nic.cfg.002	3.3
ra2.ch.015	Physical NIC Speed - Network Intensive Profile	The speed of NIC ports housed in the physical machines on which the Kubernetes Nodes run for workloads matching the Network Intensive profile must be at least 25Gbps.	infra.hw.nic.cfg.002	3.3
ra2.ch.016	Physical PCIe slots	The physical machines on which the Kubernetes Nodes run must be equipped with at least eight (8) Gen3.0 PCIe slots, each with at least eight (8) lanes.
ra2.ch.017	Immutable infrastructure	Whether physical or virtual machines are used, the Kubernetes Node must not be changed after it is instantiated. New changes to the Kubernetes Node must be implemented as new Node instances. This covers any changes from BIOS through Operating System to running processes and all associated configurations.	req.gen.cnt.02	4.3.1
ra2.ch.018	NFD	Node Feature Discovery must be used to advertise the detailed software and hardware capabilities of each node in the Kubernetes Cluster.	TBD	4.3.1

Table 4-1: Node Specifications

4.3. Kubernetes

In order for the Kubernetes components to be conformant with the Reference Architecture they must be implemented as per the following specifications:

Ref	Specification	Details	Requirement Trace	Reference Implementation Trace
ra2.k8s.001	Kubernetes Conformance	The Kubernetes distribution, product, or installer used in the implementation must be listed in the Kubernetes Distributions and Platforms document and marked (X) as conformant for the Kubernetes version defined in README.	req.gen.cnt.03	4.3.1
ra2.k8s.002	Highly available etcd	An implementation must consist of either three, five or seven nodes running the etcd service (can be colocated on the master nodes, or can run on separate nodes, but not on worker nodes).	req.gen.rsl.02 req.gen.avl.01	4.3.1
ra2.k8s.003	Highly available control plane	An implementation must consist of at least one master node per availability zone or fault domain to ensure the high availability and resilience of the Kubernetes control plane services.
ra2.k8s.012	Control plane services	A master node must run at least the following Kubernetes control plane services: kube-apiserver, kube-scheduler and kube-controller-manager.	req.gen.rsl.02, req.gen.avl.01	4.3.1
ra2.k8s.004	Highly available worker nodes	An implementation must consist of at least one worker node per availability zone or fault domain to ensure the high availability and resilience of workloads managed by Kubernetes	req.gen.rsl.01, req.gen.avl.01, req.kcm.gen.02, req.inf.com.01
ra2.k8s.005	Kubernetes API Version	In alignment with the Kubernetes version support policy, an implementation must use a Kubernetes version as per the subcomponent versions table in README.
ra2.k8s.006	NUMA Support	When hosting workloads matching the Network Intensive profile, the TopologyManager and CPUManager feature gates must be enabled and configured on the kubelet (note, TopologyManager is enabled by default in Kubernetes v1.18 and later, with CPUManager enabled by default in Kubernetes v1.10 and later). --feature-gates="...,TopologyManager=true,CPUManager=true" --topology-manager-policy=single-numa-node --cpu-manager-policy=static	e.cap.007 infra.com.cfg.002 infra.hw.cpu.cfg.003
ra2.k8s.007	DevicePlugins Feature Gate	When hosting workloads matching the Network Intensive profile, the DevicePlugins feature gate must be enabled (note, this is enabled by default in Kubernetes v1.10 or later). --feature-gates="...,DevicePlugins=true,..."	Various, e.g. e.cap.013	4.3.1
ra2.k8s.008	System Resource Reservations	To avoid resource starvation issues on nodes, the implementation of the architecture must reserve compute resources for system daemons and Kubernetes system daemons such as kubelet, container runtime, etc. Use the following kubelet flags: --reserved-cpus=[a-z], using two of a-z to reserve 2 SMT threads.	i.cap.014
ra2.k8s.009	CPU Pinning	When hosting workloads matching the Network Intensive profile, in order to support CPU Pinning, the kubelet must be started with the --cpu-manager-policy=static option. (Note, only containers in Guaranteed pods - where CPU resource requests and limits are identical - and configured with positive-integer CPU requests will take advantage of this. All other Pods will run on CPUs in the remaining shared pool.)	infra.com.cfg.003
ra2.k8s.010	IPv6DualStack	To support IPv6 and IPv4, the IPv6DualStack feature gate must be enabled on various components (requires Kubernetes v1.16 or later). kube-apiserver: --feature-gates="IPv6DualStack=true". kube-controller-manager: --feature-gates="IPv6DualStack=true" --cluster-cidr=<IPv4 CIDR>,<IPv6 CIDR> --service-cluster-ip-range=<IPv4 CIDR>,<IPv6 CIDR> --node-cidr-mask-size-ipv4 ¦ --node-cidr-mask-size-ipv6 defaults to /24 for IPv4 and /64 for IPv6. kubelet: --feature-gates="IPv6DualStack=true". kube-proxy: --cluster-cidr=<IPv4 CIDR>,<IPv6 CIDR> --feature-gates="IPv6DualStack=true"	req.inf.ntw.04
ra2.k8s.011	Anuket profile labels	To clearly identify which worker nodes are compliant with the different profiles defined by Anuket the worker nodes must be labelled according to the following pattern: an anuket.io/profile/basic label must be set to true on the worker node if it can fulfil the requirements of the basic profile and an anuket.io/profile/network-intensive label must be set to true on the worker node if it can fulfil the requirements of the network intensive profile. The requirements for both profiles can be found in chapter 2
ra2.k8s.012	Kubernetes APIs	Kubernetes Alpha API are recommended only for testing, therefore all Alpha APIs must be disabled.
ra2.k8s.013	Kubernetes APIs	Backward compatibility of all supported GA APIs of Kubernetes must be supported.
ra2.k8s.014	Security Groups	Kubernetes must support NetworkPolicy feature.
ra2.k8s.015	Publishing Services (ServiceTypes)	Kubernetes must support LoadBalancer Publishing Service (ServiceTypes).
ra2.k8s.016	Publishing Services (ServiceTypes)	Kubernetes must support Ingress.
ra2.k8s.017	Publishing Services (ServiceTypes)	Kubernetes should support NodePort Publishing Service (ServiceTypes).	req.inf.ntw.17
ra2.k8s.018	Publishing Services (ServiceTypes)	Kubernetes should support ExternalName Publishing Service (ServiceTypes).
ra2.k8s.019	Kubernetes APIs	Kubernetes Beta APIs must be supported only when a stable GA of the same version doesn’t exist.	req.int.api.04

Table 4-2: Kubernetes Specifications

4.4. Container runtimes

Ref	Specification	Details	Requirement Trace	Reference Implementation Trace
ra2.crt.001	Conformance with OCI 1.0 runtime spec	The container runtime must be implemented as per the OCI 1.0 (Open Container Initiative 1.0) specification.	req.gen.ost.01	4.3.1
ra2.crt.002	Kubernetes Container Runtime Interface (CRI)	The Kubernetes container runtime must be implemented as per the Kubernetes Container Runtime Interface (CRI)	req.gen.ost.01	4.3.1

Table 4-3: Container Runtime Specifications

4.5. Networking solutions

In order for the networking solution(s) to be conformant with the Reference Architecture they must be implemented as per the following specifications:

Ref	Specification	Details	Requirement Trace	Reference Implementation Trace
ra2.ntw.001	Centralised network administration	The networking solution deployed within the implementation must be administered through the Kubernetes API using native Kubernetes API resources and objects, or Custom Resources.	req.inf.ntw.03	4.3.1
ra2.ntw.002	Default Pod Network - CNI	The networking solution deployed within the implementation must use a CNI-conformant Network Plugin for the Default Pod Network, as the alternative (kubenet) does not support cross-node networking or Network Policies.	req.gen.ost.01, req.inf.ntw.08	4.3.1
ra2.ntw.003	Multiple connection points	The networking solution deployed within the implementation must support the capability to connect at least FIVE connection points to each Pod, which are additional to the default connection point managed by the default Pod network CNI plugin.	e.cap.004	4.3.1
ra2.ntw.004	Multiple connection points presentation	The networking solution deployed within the implementation must ensure that all additional non-default connection points are requested by Pods using standard Kubernetes resource scheduling mechanisms such as annotations or container resource requests and limits.	req.inf.ntw.03	4.3.1
ra2.ntw.005	Multiplexer/meta-plugin	The networking solution deployed within the implementation may use a multiplexer/meta-plugin.	req.inf.ntw.06, req.inf.ntw.07	4.3.1
ra2.ntw.006	Multiplexer/meta-plugin CNI Conformance	If used, the selected multiplexer/meta-plugin must integrate with the Kubernetes control plane via CNI.	req.gen.ost.01	4.3.1
ra2.ntw.007	Multiplexer/meta-plugin CNI Plugins	If used, the selected multiplexer/meta-plugin must support the use of multiple CNI-conformant Network Plugins.	req.gen.ost.01, req.inf.ntw.06	4.3.1
ra2.ntw.008	SR-IOV Device Plugin for Network Intensive	When hosting workloads that match the Network Intensive profile and require SR-IOV acceleration, a Device Plugin for SR-IOV must be used to configure the SR-IOV devices and advertise them to the kubelet.	e.cap.013	4.3.1
ra2.ntw.009	Multiple connection points with multiplexer/meta-plugin	When a multiplexer/meta-plugin is used, the additional non-default connection points must be managed by a CNI-conformant Network Plugin.	req.gen.ost.01	4.3.1
ra2.ntw.010	User plane networking	When hosting workloads matching the Network Intensive profile, CNI network plugins that support the use of DPDK, VPP, and/or SR-IOV must be deployed as part of the networking solution.	infra.net.acc.cfg.001	4.3.1
ra2.ntw.011	NATless connectivity	When hosting workloads that require source and destination IP addresses to be preserved in the traffic headers, a NATless CNI plugin that exposes the pod IP directly to the external networks (e.g. Calico, MACVLAN or IPVLAN CNI plugins) must be used.	req.inf.ntw.14
ra2.ntw.012	Device Plugins	When hosting workloads matching the Network Intensive profile that require the use of FPGA, SR-IOV or other Acceleration Hardware, a Device Plugin for that FPGA or Acceleration Hardware must be used.	e.cap.016, e.cap.013	4.3.1
ra2.ntw.013	Dual stack CNI	The networking solution deployed within the implementation must use a CNI-conformant Network Plugin that is able to support dual-stack IPv4/IPv6 networking.	req.inf.ntw.04
ra2.ntw.014	Security Groups	The networking solution deployed within the implementation must support network policies.	infra.net.cfg.004
ra2.ntw.015	IPAM plugin for multiplexer	When a multiplexer/meta-plugin is used, a CNI-conformant IPAM Network Plugin must be installed to allocate IP addresses for secondary network interfaces across all nodes of the cluster.	req.inf.ntw.10

Table 4-4: Networking Solution Specifications

4.6. Storage components

In order for the storage solutions to be conformant with the Reference Architecture they must be implemented as per the following specifications:

Ref	Specification	Details	Requirement Trace	Reference Implementation Trace
ra2.stg.001	Ephemeral Storage	An implementation must support ephemeral storage, for the unpacked container images to be stored and executed from, as a directory in the filesystem on the worker node on which the container is running. See the Container runtimes section above for more information on how this meets the requirement for ephemeral storage for containers.
ra2.stg.002	Kubernetes Volumes	An implementation may attach additional storage to containers using Kubernetes Volumes.
ra2.stg.003	Kubernetes Volumes	An implementation may use Volume Plugins (see ra2.stg.005 below) to allow the use of a storage protocol (e.g., iSCSI, NFS) or management API (e.g., Cinder, EBS) for the attaching and mounting of storage into a Pod.
ra2.stg.004	Persistent Volumes	An implementation may support Kubernetes Persistent Volumes (PV) to provide persistent storage for Pods. Persistent Volumes exist independent of the lifecycle of containers and/or pods.	req.inf.stg.01
ra2.stg.005	Storage Volume Types	An implementation must support the following Volume types: emptyDir, ConfigMap, Secret and PersistentVolumeClaim. Other Volume plugins may be supported to allow for the use of a range of backend storage systems.
ra2.stg.006	Container Storage Interface (CSI)	An implementation may support the Container Storage Interface (CSI), an Out-of-tree plugin. In order to support CSI, the feature gates CSIDriverRegistry and CSINodeInfo must be enabled. The implementation must use a CSI driver (a full list of CSI drivers can be found here). An implementation may support ephemeral storage through a CSI-compatible volume plugin in which case the CSIInlineVolume feature gate must be enabled. An implementation may support Persistent Volumes through a CSI-compatible volume plugin in which case the CSIPersistentVolume feature gate must be enabled.
ra2.stg.007		An implementation should use Kubernetes Storage Classes to support automation and the separation of concerns between providers of a service and consumers of the service.

Table 4-6: Storage Solution Specifications

A note on object storage:

• This Reference Architecture does not include any specifications for object storage, as this is neither a native Kubernetes object, nor something that is required by CSI drivers. Object storage is an application-level requirement that would ordinarily be provided by a highly scalable service offering rather than being something an individual Kubernetes cluster could offer.

Todo: specifications/commentary to support req.inf.stg.04 (SDS) and req.inf.stg.05 (high performance and horizontally scalable storage). Also req.sec.gen.06 (storage resource isolation), req.sec.gen.10 (CIS - if applicable) and req.sec.zon.03 (data encryption at rest).

4.7. Service meshes

Application service meshes are not in scope for the architecture. The service mesh is a dedicated infrastructure layer for handling service-to-service communication, and it is recommended to secure service-to-service communications within a cluster and to reduce the attack surface. The benefits of the service mesh framework are described in 5.4.3. In addition to securing communications, the use of a service mesh extends Kubernetes capabilities regarding observability and reliability.

Network service mesh specifications are handled in section 4.5 Networking solutions.

4.8. Kubernetes Application package manager

In order for the application package managers to be conformant with the Reference Architecture they must be implemented as per the following specifications:

Ref	Specification	Details	Requirement Trace	Reference Implementation Trace
ra2.pkg.001	API-based package management	A package manager must use the Kubernetes APIs to manage application artifacts. Cluster-side components such as Tiller are not supported.	req.int.api.02

Table 4-7: Kubernetes Application Package Manager Specifications

4.9. Kubernetes workloads

In order for the Kubernetes workloads to be conformant with the Reference Architecture they must be implemented as per the following specifications:

Ref	Specification	Details	Requirement Trace	Reference Implementation Trace
ra2.app.001	Root Parameter Group (OCI Spec)	Specifies the container’s root filesystem.	TBD	N/A
ra2.app.002	Mounts Parameter Group (OCI Spec)	Specifies additional mounts beyond root.	TBD	N/A
ra2.app.003	Process Parameter Group (OCI Spec)	Specifies the container process.	TBD	N/A
ra2.app.004	Hostname Parameter Group (OCI Spec)	Specifies the container’s hostname as seen by processes running inside the container.	TBD	N/A
ra2.app.005	User Parameter Group (OCI Spec)	User for the process is a platform-specific structure that allows specific control over which user the process runs as.	TBD	N/A
ra2.app.006	Consumption of additional, non-default connection points	The workload must request additional non-default connection points through the use of workload annotations or resource requests and limits within the container spec passed to the Kubernetes API Server.	req.int.api.01	N/A
ra2.app.007	Host Volumes	Workloads should not use hostPath volumes, as Pods with identical configuration (such as those created from a PodTemplate) may behave differently on different nodes due to different files on the nodes.	req.kcm.gen.02.	N/A
ra2.app.008	Infrastructure dependency	Workloads must not rely on the availability of the master nodes for the successful execution of their functionality (i.e. loss of the master nodes may affect non-functional behaviours such as healing and scaling, but components that are already running will continue to do so without issue).	TBD	N/A
ra2.app.009	Device plugins	Workload descriptors must use the resources advertised by the device plugins to indicate their need for an FPGA, SR-IOV or other acceleration device.	TBD	N/A
ra2.app.010	Node Feature Discovery (NFD)	Workload descriptors must use the labels advertised by Node Feature Discovery to indicate which node software of hardware features they need.	TBD	N/A

Table 4-8: Kubernetes Workload Specifications

4.10. Additional required components

This chapter should list any additional components needed to provide the services defined in Chapter 3.2 (e.g., Prometheus)