Kubernetes-to-AWS Security Bridge

AWS Kubernetes RBAC Bridge

Background

Security in an interconnected, always-on (24*7), virtualised, digital world is important. As more of our IT infrastructure moves to the Cloud, proactively seeking and blocking emerging security gaps becomes a continuous activity (BAU).

AWS and Kubernetes are leaders in the new paradigm of abstracted infrastructure – Cloud, datacentre on on-premise. Both have their own evolving security arrangements. For role-based access control (RBAC), AWS uses the IAM primarily, while Kubernetes (K8s) uses a combination of Roles and Role Bindings. The primary RBAC intersection point between these two has been the node/virtual-machine (EC2).

The Problem

In a simple world, privileges assigned to the underlying AWS node are inherited by the K8s Pods running on the node. This works perfectly when there is a one-to-one mapping between the client of K8s and the consumer of the AWS node. Specifically; the same entity owns the K8s cluster and the AWS node on which it runs. Security is intact, irrespective of the number of K8s Pods on the AWS node. However, misalignment occurs when K8s shares the same node among two or more clients – often referred to as multi-tenant mode. A potential for a security breach emerges.

Imagine a scenario in which there are three K8s Pods (A, B & C) running on a single AWS node. Each Pod runs a service that belongs to a different customer/client. Pod A belongs to client-A, Pod B belongs to client-B and Pod-C belongs to client-C. Files needed by each client are stored on S3 buckets in AWS, and each client has responsibility to arrange for their own S3 bucket. However, client-C is the only one that has managed to provision an S3 bucket at the time of deployment. Ordinarily, Pod A and B should never access the resource(s) provided strictly for Pod C. But if they do, nothing stops them! The diagram below provides a useful illustration.

K8S-AWS-RBAC_Quandary

Historically, in IAM, access privileges to the resource for Pod C will have been given to the node hosting Pods A, B and C. The EC2 node would have an Instance Profile defined, and a Role will be attached to the Instance Profile, giving it those privileges. The unexpected consequence however is that Pods A and B also inherit the privilege from the host node. Pod C’s resources would therefore be accessible to any other Pod (client) running on that node. This obviously is not acceptable for a multi-tenant K8s cluster.

Solutions

The great thing about the Open Source community is that problems are attacked, and often solved, almost as soon as they are articulated. Two open source products emerged to close this security gap: Kube2IAM (2016) and KIAM (2017). Some time later, AWS introduced a solution; “IAM for Service Accounts”. However, the AWS solution only works with their EKS service. All three solutions make it possible to control access from K8s Pods to AWS resources.

I will not discuss the AWS solution as it is proprietary and closely tied to their EKS offering. Neither will I examine KIAM as the solution has been abandoned by the developers. This leaves us with the forerunner: Kube2IAM. Kube2IAM deploys a K8s DaemonSet in the K8s cluster. By default, one Kube2IAM Pod is deployed to each worker node in the cluster. The Kube2IAM instance running on each node intercepts requests to the AWS metadata service URL (http://169.254.169.254). It then provides a response according the the IAM role assignments, as well as the annotations on the Pod calling the service. The diagram below provides a useful illustration.

AWS Kubernetes RBAC Bridge

With this simple solution by Kube2IAM, the exclusive role assignment to Pod C is respected by K8s. Deliberate or inadvertent requests by Pod A or B are blocked by Kube2IAM.

Here is how it works. When a Pod makes a request for AWS resources, it will make a call to the AWS metadata service URL. Kube2IAM hijacks the call (iptables reroute) and performs an inspection to see what the appropriate response should be. It checks if there are any appropriate RBAC annotations on the Pod making the request. If there are none, Kube2IAM serves up the default privilege set. These will be the privileges defined for the EC2 Instance Profile. However, if the Pod has a role annotation, it will be given the privileges defined in the matching AWS role.

Hands-on

In the example that follows, we will deploy two Pods; one with annotations (annotated) and another without (vanilla). We will use two AWS roles. The read-only role will have access to one S3 bucket only. The other read+write role will have read access to 2 buckets and read+write access to one bucket. The read-only role will be attached to the EC2 Instance Profile for the K8s worker node. The read+write role be standalone, but it will be extended to trust the read-only role. This sets the stage for Kube2IAM to discriminate between requests, giving read and/or write access to our Pods, as appropriate. In our example, the annotated Pod will be able to write one bucket and read two buckets, while the vanilla Pod will only be able to read one bucket.

The implementation artefacts can be downloaded from GitHub (use this link). I have put together what I think is a simple, and perhaps more explicit set of instructions below. Follow them step-by-step and you should end up with a working RBAC bridge using Kube2IAM. I guess one could write a script that automates all of these steps, but that is a task for another day, or perhaps someone else.

Process

  1. Create a policy (nettech-s3-read-only); use the file nettech-ec2-instance-profile.json for the policy definition/contents
  2. Create a role (nettech-s3-read-only); the role should refer to the policy in step #1
  3. Create an EC2 instance profile (nettech-instance-profile) for the AWS node(s); the instance profile should refer to the role you defined in step #2, forming a chain:
    nettech-instance-profile==>nettech-s3-read-only(role)==>nettech-s3-read-only(policy).
      Use the following aws-cli commands:
      aws iam create-instance-profile –instance-profile-name nettech-instance-profile
      aws iam add-role-to-instance-profile –instance-profile-name nettech-instance-profile –role-name nettech-s3-read-only
  4. Create a second read+write S3 policy and role (nettech-s3-read-write). Use the file nettech-s3-read-write.json for the policy definition/contents
  5. Extend the trust relationship on the read+write S3 role such that it can be assumed by the read-only role, forming a link:
    nettech-s3-read-write(role)<==trust==nettech-s3-read-only(role).
      In IAM console, select the read+write role
      Select the “Trust relationships” tab, and then click on the “Edit trust relationships” button
      In the new window that opens, add the contents of the file nettech-s3-read-write-trust-relationship.json to the existing definition/contents
      Make sure to update the AWS account Id (01234567890) to your own
      Click on “Update Trust Policy” to save your changes
  6. Deploy or assign a worker node in your K8s (Rancher/Kops/..) cluster
  7. Configure or update the worker node to reference the EC2 Instance Profile (nettech-instance-profile) from step #3
      aws ec2 associate-iam-instance-profile –iam-instance-profile nettech-instance-profile –instance-id xxxxxxxxxx # replace xxxx with your instance Id, or use the AWS GUI to attach it
  8. Deploy Nginx vanilla and annotated (K8s Deployments). Use the file nginx-deployment.yaml from Rancher UI or kubectl on the command line
  9. Install aws-cli in each of the Nginx instances – Use the following commands (Linux):
      apt update
      apt install curl -y
      apt install unzip -y
      curl "https://awscli.amazonaws.com/awscli-exe-linux-x86_64.zip" -o "awscliv2.zip"
      unzip ./awscliv2.zip
      ./aws/install
  10. Verify that the host node has read access to the “nettech-helm-test” bucket, according to the EC2 profile from step #3. Connect to the host node (via Rancher UI or SSH) and run the aws s3 ls command.
      aws s3 ls nettech-helm-test
  11. Verify that both Pods have read access to the “nettech-helm-test” bucket. Connect to each Pod (via Rancher UI or kubectl) and run an aws s3 ls
      aws s3 ls nettech-helm-test
  12. Create/deploy ClusterRole & ClusterRoleBinding for the service account to be used by Kube2IAM. Use the file clusterRoleAndBinding.yaml
  13. Deploy Kube2IAM (K8s DaemonSet), with debugging enabled. Use the file kube2IAM-daemonSet.yaml
  14. Connect to the host node and access the command line. Check that only one IPTables rule exists on the worker node (for AWS metadata IP). Delete any duplicates to avoid confusing errors. This may happen if you redeploy the Kube2IAM Daemonset.
      sudo iptables -t nat -S PREROUTING | grep 169.254.169.254 # list all entries
      sudo iptables -t nat -D PREROUTING -d 169.254.169.254/32 -i docker0 -p tcp -m tcp –dport 80 -j DNAT –to-destination 10.43.0.1:8282 # delete any duplicates
      NB: (docker0) is the network interface, (10.43.0.1) is the IP address of the node/host, and (8282) is the Kube2IAM port
  15. Test the Nginx instances again
    • Verify that the host node still only has read access to “nettech-helm-test”, as defined as a default in the EC2 Profile role (nettech-s3-read-only)
    • Verify that the vanilla Nginx Deployment still only has read access to “nettech-helm-test”, as defined as a default in the EC2 Profile role (nettech-s3-read-only)
    • Verify that the annotated Nginx Deployment now has read access to “lanre.k8s.dev” and “nettech-helm-test” as well as read+write access to “lanre.k8s.dev”

Conclusion

An RBAC bridge of some sort is a necessity for all multi-tenant K8s clusters running on virtualised infrastructure such as AWS, Azure, GCP and others. Kube2IAM provides an effective solution for the AWS platform. This article identifies the issue that Kube2IAM resolves and shows a very simple, sandbox implementation. The article should serve as quick-start guide that is easy to grasp and quick to implement.

We live in a rapidly evolving technology environment. Kube2IAM has set a very sound foundation, but as always, there is always room for improvement; and I say that with all humility and respect for the developers. KIAM came up with a cacheing service to reduce latency and improve scalability, unfortunately, that solution is no longer being evolved. One would like to see similar functionality in Kube2IAM. One other improvement would be to move the annotations out of the Pod and into K8s roles. The preference being roles defined outside the namespace of the beneficiary Pod. This will reduce the attack surface for malicious code that might attempt a brute-force attack to find AWS roles that can be exploited.

Many thanks to Jerome Touffe-Blin @jtblin and his team for creating this ever-so-useful open-souce utility.



Oyewole, Olanrewaju J (Mr.)
Internet Technologies Ltd.
lanre@net-technologies.com
www.net-technologies.com

Cloud Migration and Availability Index

Infrastructure Migration

The Cloud Transition

The Cloud may not be in-your-face.  But it is pervasive, and gradually taking over many aspects of our traditional IT systems. Companies are not yet making wholesale transitions from existing data-centres and on-premise assets to Cloud. However, when infrastructure reviews occur, whether to upgrade or add new resources, the Cloud beckons. Questions about total cost of ownership (TCO), scalability, time-to-market, etc will influence decision makers.  For each one of these, the Cloud offers a compelling alternative. It is likely that in the next two decades, only a few companies will still maintain their infrastructure on premise.

The Status Quo - On-premise Deployment Design
Let us assume then that ACME plc has made a decision. Business has been persuaded, either by hype or fundamentals, that the Cloud is the strategic target. Architectural leadership has been mobilised and a decision taken to draw up a roadmap for Cloud adoption. What next? In this article, we look at four primary considerations that architects must carefully examine when migrating to the Cloud. These are: sizing, failover, scaling and access. Everything else builds on the foundation that is synthesised from these four dimensions.

Sizing: What Specification of Infrastructure Should be Provisioned

Statistics are invaluable. Node sizing should be empathetic to existing use profile. It may be okay to guess at first, but it saves so much time to know in advance. For each Cloud instance, the node(s) provisioned should be selected to meet latency and throughput required to support 120% of anticipated production load. The sizing could be either singular or plural. Singular, as in one node with enough capacity to bear all load; or plural, i.e. a number of nodes that can, between them, satisfy demand. But the baseline should exceed the present need.

Resizing in the Cloud may be quick and easy, but the decision making might not be so. If in doubt, over-provision. It is easy to downsize later, and the organisation avoids the risk of loss of business due to performance or availability problems. Default sizing is simple, i.e. geography localised and singular. But there could be exceptional scenarios where geographic distribution must be combined with plural sizing. More about that later.

Failover: How is System Failure Mitigated

Given proper sizing, as above, the next dimension to consider is failure and recovery. If or when a properly sized machine fails; what happens next? Let’s take the simple approach first and we will revisit this later. There should be a distribution of node(s) across Cloud locations, so that the failure of one node does not result in service unavailability. Service recovery should occur in a different Cloud location. This reduces the likelihood of contagion from the original failure location while maintaining service continuity. An interesting aspect of failure management is implicit resilience, i.e. what measure of interuption can our infrastructure handle?

The complement of the nodes that provide a service across Cloud location(s) is a resource group. The group resilience is the count of simultaneous failures that can be managed while maintaining SLAs. The higher the count, the larger the number of nodes and Cloud locations involved. Resiliency has a price tag though. More machines (virtual) will multiply cost and increase the percentage of idle/redundant resources in the Cloud platform.

Scaling: How are Additional Resources Provisioned

As resource demand grows organically, or due to unexpected spikes, infrastructure should respond, automagically! Traditionally, scaling was a bureaucractic and technical journey. With Cloud, scaling is merely a change of configuration. Where singular sizing has been used, another node of the same size could be added. This is horizontal scaling. Adding more nodes to singular sized nodes would multiply capacity. It is linear, but there is no guarantee of commensurate increase in demand or resource usage. There is an alternative design that is more efficient: programmatic vertical scaling. A simple algorithm can be applied to automatically scale resources; up or down, by a fraction rather than a multiple.

Cloud platforms record a raft of events about the resources deployed. Customers can tap in to these events to scale in response to demand. On AWS, CloudWatch alarms can trigger a Lambda function, which in turn effects a rolling upgrade on EC2 nodes; upscaling node size before autoscaling. By leveraging statistics for baseline sizing and monitoring demand, we can guarantee day zero availability and decent response in infrastructure provisioning. Increasing capacity as demand grows and shrinking it if or when spikes even out.

Access: How do Clients Connect to Cloud Services

The fourth dimension is access. As on-premise, so also with Cloud. There is no value in having resources that are locked away from everyone and everything. Our clients need access to our Cloud based services, so also partners involved in our service chain. Unless we are migrating all at once, it is likely that we will also need access to some on-premise infrastructure. Our design must provide the access paths and levels, as well as the constraints that keep authorised clients within band and everyone else out. To achieve this we would use such facilities as the Virtual Private Network (VPN), the load balancer, firewalls and others. Beyond the basics of who’s in and who’s out though, there is a service that we must provide to clients and partners.

The key here is to be simple and unobtrusive; placing minimal burdens on clients, partners and our applications/services.

By default we would use load balancers to decouple clients from service providers. Cloud load-balancers spread requests among available service providers. They are not geography specific and simplify access and security for clients and service provider.  Our Cloud landscape is elegant and uncomplicated, with singular entry points for each service.  One consideration could however force radical change to this default: Geographic Affinity (GA).  Geographic affinity is a requirement to pin clients to a specific physical/geographical service provider.  It could be zonal or regional. GA is often driven by regulatory, localisation, performance or security concerns.

But some GA drivers can be conflicting. For example, performance (latency sensitive applications) might be a problem where localisation (regional) is required. Invariably, GA tilts our architecture towards plurality of nodes and complications in managing performance and synchronisation of state. Architects must balance, sometimes conflicting, needs to avoid creating virtual silos in the Cloud.

Cloud Deployment Design
Cloud Chaos

The Availability Index

So far we have been working forwards from an existing status quo to a target architecture. We have also adopted an exclusively technical perspective. What would be better is to take a business perspective. To approach our context top down. We should ask: what infrastructure is needed to support our business vision, now and into the near future? What level of availability is enough to provide service that exceeds client needs. In asking these questions, we encounter a new concept: “the Granularity of Perception”. This can be described as the number of microseconds, milliseconds, seconds, minutes, or more that impacts our service(s), as perceived by clients. Simply put: how slowly can we blink before our clients start to notice that our eyes have moved. As this number (granularity) increases, the required level of availability decreases. The table below provides a rough guide, with descriptions.

Availability Index Description
1 Cluster enabled, auto recovery, no fail 24×7, latency intolerant, high-frequency, geography affinity
3 Cluster enabled, auto recovery, no fail 24×7, latency intolerant, medium frequency
5 Cluster enabled, auto failover, business hours, latency tolerant, low frequency
7 Non clustered, manual failover, business hours, latency tolerant, low frequency

The goal of architects should be to design a Cloud platform that delivers a granularity that is finer than the perception of clients.  Using the table above as a guide, architects should play out scenarios with the service portfolio against each index.  Starting with the least to the highest.  Once the required availability index is determined, it should be relatively easy to identify the dimensions to support it.

Conclusion

As organisations embark on the journey of digital transformation, one early change is often Cloud adoption. This is because the Cloud provides a catalysing medium in which many solutions are easier and quicker to provision.  In moving from on-premise/data-centre resources to the Cloud, architects must resist the temptation to simply lift-and-shift.  Rather, the digital transformation journey should re-examine the fitness-for-purpose of existing solutions, platforms and infrastructure. There is a famous quote by Jack Welch, former CEO of General Electric. He said, If the rate of change on the outside exceeds the rate of change on the inside, then the end is near.. In a rapidly evolving globalised economy, business agility is becoming a precondition for survival.

The availability index is a simple, logical, technology-agnostic technique for conceptual reasoning about a Cloud landscape.  Determination of the availability index helps to reveal shared profiles for similar subsystems.  The profiles are logical and help estimate the resources required to support a genre of subsystem.  Each logical profile can then be mapped to specific Cloud infrastructure and captured as customisable templates.  The logical profiles provide architects with a starting point for solution designs.  The infrastructure templates serve as a baseline for DevOps teams.  Each artefact is likely to go through a number of evolutions.  However, it is vital that both views are kept in sync at all times.

Organisations that leverage this approach will see a marked improvement in the consistency of infrastructure designs.  Other benefits include faster turnaround of solutions, and systems that balance technical capability with business needs and aspirations. Architecture teams that leverage the availability index position their organisations for superior agility and competitiveness in the global economy.


Oyewole, Olanrewaju J (Mr.)
Internet Technologies Ltd.
lanre@net-technologies.com
www.net-technologies.com

Securing our Hybrid Cloud

Cloud Infrastructure Security

Cloud Infrastructure SecurityIn a previous article, “One thousand servers, start with one click”, I described the design and implementation of a simple hybrid-Cloud infrastructure.  The view was from a high level, and I intend, God willing, to delve into the detail in a later instalment.  Before that though, I wanted to touch on the subject of hybrid Cloud security, briefly.

Having deployed resources in a private Cloud and on-premise, certain precautions should be taken to secure the perimeter and the green zone inside the two networks – Cloud and premise.  The diagram above paints the big picture.  It shows how, based on AWS security recommendations, minimal access/privilege is granted to each resource in the networks.  The focus here is on machine access, which is about the lowest level of security.  I will not delve into AWS policies, VPN configuration or on-premise firewall rules, as these are not black-and-white and the detail involved does not fit in with the goal for this article.

Here goes!  Reading the diagram and translating to words:

  1. It is convenient to use the Internet Gateway (public router) of your Cloud provider during development.  Once you are done with prototyping and debugging, it should be disabled or removed.  Switch to a NAT gateway (egress only router) instead.  Your servers can still connect to the outside world for patches, updates, etc. but you can control what sites are accessible from your firewall.  Switching to a NAT gateway also means that opportunist snoopers are kept firmly out.
  2. Open up port 443 for your Kubernetes (K8S) master node(s) and close all others – your cluster can even survive without the master node, so don’t by shy, lock it down.  Should the need arise, it is easy to temporarily change Cloud and premise firewall rules to open up port 22 (SSH) or others to investigate or resolve issues.  Master nodes in K8S master subnet should have access to all subnets within the cluster, but this does not include known service ports for the servers or databases.
  3. While your ESB/EI will have several reusable/shared artefacts, the only one that are of interest to your clients (partners) are the API and PROXY services.  For each one of these services, a Swagger definition should be created and imported into the API Manager (APIM).  All clients should gain access to the ESB/EI only through the interfaces defined in the APIM, which can be constrained by security policies and monitored for analytics.  Therefore, the known service access ports should be open to clients on the APIM, and as with the K8S master, all other ports should be locked down.
  4. Access to the known service ports on the ESB/EI should be limited to the APIM subnet(s) only, all other ports should be closed.
  5. The Jenkins CI/CD containers are also deployed to the same nodes as the ESB/EI servers, but they fall under different constraints.  Ideally, the Jenkins server should be closed off completely to access from clients.  It can be correctly configured to automatically run scheduled and ad-hoc jobs without supervision.  If this is a challenge, the service access port should be kept open, but only to access from within the VPN, ideally, a jump-box.
  6. Access to the cluster databases should be limited to the APIM and ESB/EI subnets only, and further restricted to known service ports – 3306 or other configured port.
  7. Access to the cluster NFS should be limited to the APIM, ESB/EI, and K8S-master subnets only, and further restricted to known service ports – 111, 1110, 2049, etc., or others as configured.
  8. On-premise firewall rules should be configured to allow access to SFTP, database, application and web-servers from the ESB/EI server using their private IP addresses over the VPN.
  9. Wherever possible, all ingress traffic to the private Cloud should flow through the on-premise firewall and the VPN.  One great benefit of this approach is that it limits exposure; there are fewer gateways to defend.  There are costs though.  Firstly, higher latencies are incurred for circuitous routing via the VPN rather than direct/faster routing through Cloud-provider gateways.  Other costs include increased bandwidth usage on the VPN, additional load on DNS servers, maintenance of NAT records, and DNS synchronisation of dynamic changes to nodes within the cluster.
  10. ADDENDUM: Except for SFTP between the ESB/EI server and on-premise SFTP servers, SSH/port-22 access should be disabled.  The Cloud infrastructure should be an on-demand, code-driven, pre-packaged environment; created and destroyed as and when needed.

And that’s all folks! Once again, this is not an exhaustive coverage on all the aspects of security required for this hybrid-Cloud.  It is rather a quick run-through of the foundational provisions.  The aim being to identify a few key provision that can be deployed very quickly and guarantee a good level of protection on day one.  All of this builds on a principle adopted from AWS best practise. The principle states that AWS is responsible for the security of the Cloud while end-users are responsible for security in the Cloud. The end-user responsibility of Cloud security begins with another principle: access by least privileges. This means that for a given task, the minimum privileges should be granted to the most restricted community to which one or more parties (man or machine) is granted membership.


Oyewole, Olanrewaju J (Mr.)
Internet Technologies Ltd.
lanre@net-technologies.com
www.net-technologies.com
Mobile: +44 793 920 3120

1000 Cloud Servers: Start With One Click

[Hybrid] Cloud Infrastructure

Cloud and Open Source

The arrival of Cloud providers and Infrastructure-as-a-Service (IaaS) has opened up options and possibilities for solution architects.  Our company is working with a client on a major transformation initiative. Leveraging Cloud IaaS and open-source integration platforms, together we have explored options, built competence, and delivered incremental solutions while keeping costs to a minimum.  Without Cloud IaaS and open-source this freedom of expression in solution architecture would have been impossible.  Just imagine justifying a multi-tier, multi-server solution to the CFO when one of the key drivers has been cost control!

 

The Basic Idea

In its most primitive expression, our client wanted a public Application Program Interface (API) layer to abstract access to an Integration layer, which in turn connected with all their internal repositories and partner systems to provide services.  The image that follows provides an illustration.  It appears quite straightforward and simple.
DraftInfrastructure

The API layer provides a simple Representation State Transfer (REST) interface as well as security.  It also maintains logs that can be analysed for insights into client behaviour and the usage/performance of services.  The Integration layer serves as an Enterprise Service Bus (ESB), connecting to databases, FTP and/or file servers, as well as internal and partner web services.  In addition, it manages the interactions between all participating systems in the enterprise and ensures that valuable services are made available to the API layer.

 

Enter Cloud (AWS/Azure) and Open Source (WSO2)

The traditional route would have been to procure/secure access to servers in a data-centre or in-house server-room and buy licenses from a vendor.  That would have meant a lead time of several weeks or months, to negotiate the price of licenses and consultancy, arrange for servers and networking, and to secure and disburse requisite financing.  With Cloud and open-source software, upfront costs were near-zero.  The greatest resource demand was the effort required to architect the Cloud infrastructure and to create the code to build, populate and operate it.

 

Building the Foundation

There were many options for building the networking and computing instances.  We chose Kubernetes.  Kubernetes is well established and provides abstractions that make it easy to switch Cloud providers.  Using the analogy of a house for illustration; Kubernetes builds the shell of the house, setting up the rooms, corridors, and spaces for windows and doors.  It keeps a record of each component, and notifies all other components if there has been a change to any one of them.  In our use case, Kubernetes creates a private network in the Cloud (cluster), adds compute-instances, load-balancers, firewalls, subnets, DHCP servers, DNS servers, etc. Kubernetes also provides a dynamic registry of all these components that is kept up to date with any changes in real time.

 

The First Change: Redundancy

In the past, vertical scaling with large singleton servers was standard.  These days, horizontal scaling with smaller machines (compute instances) that adjust to changing needs is the norm.  This new approach also provides fail safety.  If one machine fails, there will be other(s) to take up the load. Fortunately this is a core feature of Kubernetes.  The cluster monitors itself to ensure that all the declared components are kept alive.  If/when a component fails, the management services within the cluster ensure that it is replaced with an identical component.  For this reason, rather than have one instance of each component, two or more are created and maintained.

 

The Second Change: Dynamic Delivery

We could have chosen to install all of our technology stack (software) on each compute instance on creation.  That would be slow though, and it could mean that the instances would need to be restarted or swapped-out more often as memory and/or disk performance degrade.  Instead of that, we used Docker to build Containers that are delivered to the instances. The Docker Containers borrow memory, CPU and other resources from the compute instance at the point of deployment.  Containers can be swapped in and out, and multiple Containers can be run on the same compute instance.  When a Container is stopped or removed, the block of borrowed resources are returned to the compute instance.   A Container can be likened to a prefabricated bathroom; it is built offsite and plumbed in at delivery.  Unlike a technology stack that is built from scratch over minutes or hours, a Container is usually ready for access within a few seconds/minutes of its deployment.
VMInfrastructure

 

Implicit Change: Clustering

When more than one instance of a genre component is running at the same time, the complement of all is referred to as a cluster.  Components running in a cluster have peculiar needs; one of which is the sharing of state (status).  State is a snapshot of the world from the computer’s perspective.  In a cluster, all component instances must share the same configuration and operate on the same data always.  To facilitate this, we introduced two repositories.  A Network File System (NFS) for sharing configuration details, and a database for sharing operational data.  Kubernetes does not create these resources.  We used Terraform, another abstraction technology, to create the NFS and a replicated multi-zone database.  Terraform creates these in two private subnets within the private network created by Kubernetes.  Having created the NFS and database though, there was a need to configure and populate them with necessary data upfront.  While Terraform could be manhandled to achieve this, it is not really it’s raison detre.  Another tool is more suited to operating at a fine detail on remote machines: Ansible.  We created Ansible playbooks to configure users, directories and files on the NFS and to create instances, users and tables in the database.

 

Implicit Change: Discovery

The next challenge that our architecture threw up was discovery.  Within our network, there was an API layer and an EI layer.  In each of these layers, there could be several compute instances, and on each compute instance there could be one or more Docker Containers.  Beyond the API and the EI layers, there were also databases and a network file system.  How would clients or our platform gain access to our components, and how would the machines in one layer find those in another layer?  The Kubernetes configuration includes ClusterIP services that provide a single DNS name that resolves to all the compute instances for a given component.  For example, any API Container could be reached using a DNS name such as: cnt.api.example.com.  Clients of our platform could therefore use a DNS name to connect to an API Container, and any API Container could likewise use a single DNS name to communicate with a Docker Container in the EI layer.  Both the API and EI layers use a DNS name to communicate with the NFS and the database.  The IP address of the underlying components might change, but the DNS name is constant for the life of the platform, giving ease of discovery and stability.

 

Tying it all Up

It is all well and good that the components in the Cloud are in place and can talk with each other.  However, most of our operational systems are still on-premise; how do we join things up?  We created a VPN connection between the network in the Cloud and our on-premise network and set up Firewall rules to allow access to and from the Cloud.  The ClusterIP services were also revised to permanently maintain two static IP addresses.  This makes it easy to integrate them with on-premise DNS servers and thereby open them up to access from clients.  Below is an image that shows what it all looks like.
[Hybrid] Cloud Infrastructure

The Thousand Servers

All of these components, configurations, and customisations have been documented as scripts, configuration files and resources. The creation of a Cloud environment is reduced to running a script with two parameters: the name of the environment and the desired Cloud subnet.  By integrating this script into an on-premise CI/CD server, it is now possible to spin up as many Cloud environments as we like; at the click of a button.

All this is quite high-level and simplified; in the next instalment (One Thousand Servers: Start with a Script), I intend to drop down to eye-level and throw up some of the details of how we implemented all of this.  Watch this space for the activation of link above.


Oyewole, Olanrewaju J (Mr.)
Internet Technologies Ltd.
lanre@net-technologies.com
www.net-technologies.com
Mobile: +44 793 920 3120