Tag: microservices

Microservice Adoption

I worry that companies are deconstructing their monolithic applications into microservices because it’s trendy. In fact, there are places where microservices don’t make sense but rather impart additional complexity to an application that is not enhanced by the benefits of microservices. While some challenges to microservice adoption are transient or can be addressed through business decisions … some are fundamental aspects of the architecture.

Microservices are (relatively) new. Whereas a company that has built and run many monolithic applications has network, hypervisor, OS, deployment, and application experts … unless the company hires in a container orchestration / API gateway expert or brings in a consulting team (real world experience has been “learning it” was left up to employee initiative and the global archive of IT knowledge that is the Internet), there isn’t a deep knowledge base to support the framework. Not an insurmountable problem, and frankly no different than how virtualization was introduced — there weren’t hypervisor experts at the time, no one really understood sizing/scaling intricacies. It was learned, but the first 6-12 months were rough. High availability applications were physically designed to withstand failure. Our data centre has two unique circuits run to each rack – and dual power-supply servers are plugged into both the “A” and “B” circuits. Same with network – there’s a team that goes through two different switches. In switching to VM’s … we had to identify where this server runs (i.e. what is it’s host)? Is every component of a redundant system co-located on a single hypervisor or in SAN-booted VMs are they stored on a single SAN frame? Microservices will have a similar challenge — where is it running, can the service as a whole survive a fault? How do we recover from a major data centre failure?

Some of the places where I see microservices making development and operations more complex can be eliminated by business policies. Allowing individual service teams to dictate their own development language can reduce mobility between teams — the Java guru for service A will spend time researching the c# equiv if they move over work on service B. And while it is possible to publish a general coding standard that covers all languages (how variables will be named, what comment blocks should look like, etc) there are nuances to each language that make a shared standard impossible. Using multiple development languages limits employee mobility, and it also reduces a company’s ability to shift employees around to cover temporary resource shortfalls. While planned absences can be accounted for when selecting work for the next cycle, emergencies happen.

Breaking an application into small component services can create challenges in troubleshooting issues. There may be few who have an end-to-end understanding of the application. Where the monolithic application X getting munged information means the development team for App X needs to debug and sort the issue … ten interacting microservices can mean ten groups saying nothing’s wrong on their side and it’s everyone else’s problem. I’ve seen that occur frequently in infrastructure support — app guys says it is the server, server guy says it’s the hypervisor, hypervisor guy says it’s the SAN, SAN guy says it’s all good and someone should check with the network guys to see how those load balancers are doing.

Fundamentally, microservice architecture introduces additional components to run the application — the API gateway and container orchestration are functions that simply don’t exist in a monolithic application. These services themselves, as well as the supporting technologies that allow these services to function, add additional complexity.

As an example, the networking configuration behind making microservices available are not, in my experience, something with which developers are familiar. This is not a problem when dev teams require out-of-box functionality and said functionality is working properly. I became involved with container orchestration system because a friend’s dev team encountered failures where kube-proxy did not create the required iptables rules — a quick and easy thing for a Linux/Unix admin to identify and troubleshoot, but not something that concerned application developers in monolithic deployments. Since then, the dev team sought to use multiple network interfaces and the Kubernetes CNI plugin did not support that feature.

For an application where individual components have different utilization rates, microservice architecture makes sense. Thinking about a company that runs a major promotion. There will (hopefully) be a flood of customers browsing the web site. The components that handle browsing and search functions need to grow significantly. The component that handles existing user authentication, new user registration, customer checkout, inventory update, and shipping quote generation components don’t need to scale at the same level — only a fraction of the web traffic will actually convert to sales. So there’s no need to spin up new hosts in the web farm to handle users browsing product information.

For an application where individual components require frequent updates, microservice architecture makes sense. Is there a component that suffers frequent failures where having a pool of microservices available would increase the application’s uptime?

Kubernetes Sandbox With Minikube

A scaled down sandbox can be used to gain experience with the applications and techniques used to deploy containerized applications and microservices. This sandbox will be built on a Windows 10 laptop, but the same components can be run on Linux variants.


Verify Virtualization is enabled:

Open Task Manager (taskman.exe) and ensure the virtualization extensions have been enabled.

If virtualization is disabled, boot into the system config (start menu => settings => update & security => recovery, click “Restart now” under “Advanced startup”)

Uninstall the Windows OpenSSH client

Click ‘Start’ and type “Manage optional features” – within the installed feature list, find “OpenSSH Client”. If present, remove it.

Enable Hyper-V

Enable the Hyper-V Windows feature (Control Panel => Programs => Programs and Features, “Turn Windows features on or off” and check both Hyper-V components).

Add Virtual Switch To Hyper-V

In the Hyper-V Manager, open the “Virtual Switch Manager”. Create a new External virtual switch. Record the name used for your new virtual switch.


Install Minikube

View https://storage.googleapis.com/kubernetes-release/release/stable.txt and record the version number. The current stable release version is v1.11.1

Modify the URL http://storage.googleapis.com/kubernetes-release/release/v#.##.#/bin/windows/amd64/kubectl.exe to use the current stable release version. Current URL is http://storage.googleapis.com/kubernetes-release/release/v1.11.1/bin/windows/amd64/kubectl.exe

Create a folder %ProgramFiles%\Minikube and add this folder to your PATH variable.

Download kubectl.exe from the current release URL to %ProgramFiles%\Minikube

Download the current Minikube release from https://github.com/kubernetes/minikube/releases (scroll down to the “Distribution” section, locate the Windows/amd64 link, and save that binary as %ProgramFiles%\Minikube\minikube.exe). ** v0.28.1 was completely non-functional for me (and errors were related to existing issues on the minikube GitHub site) so I used v0.27.0

Verify both are functional. From a command prompt (run as administrator) or Powershell (again run as administrator), run “kubectl version” and verify the output includes a client version

Run “minikube get-k8s-versions” and verify there is output.

Configure the Minikube VM using the Hyper-V driver and switch you created earlier.

minikube start –vm-driver hyperv –hyperv-virtual-switch “Minikube Switch” –alsologtostderr

Once everything has started, “kubectl version” will report both a client and server version.

You can use “minikube ip” to ascertain the IP address of your cluster

If the cluster services fail to start, there are a few log locations.

Run “minikube logs” to see the log information from the minikube virtual machine

Use “kubectl get pods –all-namespaces” to determine which component(s) fail, then use “kubectl logs -f name -n kube-system” to review logs to determine why the component failed to start.

If you need to connect to the minikube Hyper-V VM, the username is docker and the password is tcuser – you can ssh into the host or connect to the console through the Hyper-V Manager.

Before the management interface comes online, you can use view the status of the containers using the docker command line utilities on the minikube VM.

$ docker ps

CONTAINER ID        IMAGE                        COMMAND                  CREATED              STATUS              PORTS               NAMES

7d8d66b5e465        af20925d51a3                 “kube-apiserver –ad…”   About a minute ago   Up About a minute                       k8s_kube-apiserver_kube-apiserver-minikube_kube-system_0f6076ada4273000c4b2f846f250f3f7_3

bb4be8d267cb        52920ad46f5b                 “etcd –advertise-cl…”   7 minutes ago        Up 7 minutes                            k8s_etcd_etcd-minikube_kube-system_0199781185b49d6ff5624b06273532ab_0

d6be5d6ae360        9c16409588eb                 “/opt/kube-addons.sh”    7 minutes ago        Up 7 minutes                            k8s_kube-addon-manager_kube-addon-manager-minikube_kube-system_3afaf06535cc3b85be93c31632b765da_1

b5ddf5d1ff11        ad86dbed1555                 “kube-controller-man…”   7 minutes ago        Up 7 minutes                            k8s_kube-controller-manager_kube-controller-manager-minikube_kube-system_d9cefa6e3dc9378ad420db8df48a9da5_0

252d382575c7        704ba848e69a                 “kube-scheduler –ku…”   7 minutes ago        Up 7 minutes                            k8s_kube-scheduler_kube-scheduler-minikube_kube-system_2acb197d598c4730e3f5b159b241a81b_0

421b2e264f9f        k8s.gcr.io/pause-amd64:3.1   “/pause”                 7 minutes ago        Up 7 minutes                            k8s_POD_kube-scheduler-minikube_kube-system_2acb197d598c4730e3f5b159b241a81b_0

85e0e2d0abab        k8s.gcr.io/pause-amd64:3.1   “/pause”                 7 minutes ago        Up 7 minutes                            k8s_POD_kube-controller-manager-minikube_kube-system_d9cefa6e3dc9378ad420db8df48a9da5_0

2028c6414573        k8s.gcr.io/pause-amd64:3.1   “/pause”                 7 minutes ago        Up 7 minutes                            k8s_POD_kube-apiserver-minikube_kube-system_0f6076ada4273000c4b2f846f250f3f7_0

663b87989216        k8s.gcr.io/pause-amd64:3.1   “/pause”                 7 minutes ago        Up 7 minutes                            k8s_POD_etcd-minikube_kube-system_0199781185b49d6ff5624b06273532ab_0

7eae09d0662b        k8s.gcr.io/pause-amd64:3.1   “/pause”                 7 minutes ago        Up 7 minutes                            k8s_POD_kube-addon-manager-minikube_kube-system_3afaf06535cc3b85be93c31632b765da_1


This allows you to view the specific logs for a container that is failing to launch

$ docker logs 0d21814d8226

Flag –admission-control has been deprecated, Use –enable-admission-plugins or –disable-admission-plugins instead. Will be removed in a future version.

Flag –insecure-port has been deprecated, This flag will be removed in a future version.

I0720 16:37:07.591352       1 server.go:135] Version: v1.10.0

I0720 16:37:07.596494       1 server.go:679] external host was not specified, using

I0720 16:37:08.555806       1 feature_gate.go:190] feature gates: map[Initializers:true]

I0720 16:37:08.565008       1 initialization.go:90] enabled Initializers feature as part of admission plugin setup

I0720 16:37:08.690234       1 plugins.go:149] Loaded 10 admission controller(s) successfully in the following order: NamespaceLifecycle,LimitRanger,ServiceAccount,NodeRestriction,DefaultTolerationSeconds,DefaultStorageClass,MutatingAdmissionWebhook,Initializers,ValidatingAdmissionWebhook,ResourceQuota.

I0720 16:37:08.717560       1 master.go:228] Using reconciler: master-count

W0720 16:37:09.383605       1 genericapiserver.go:342] Skipping API batch/v2alpha1 because it has no resources.

W0720 16:37:09.399172       1 genericapiserver.go:342] Skipping API rbac.authorization.k8s.io/v1alpha1 because it has no resources.

W0720 16:37:09.407426       1 genericapiserver.go:342] Skipping API storage.k8s.io/v1alpha1 because it has no resources.

W0720 16:37:09.445491       1 genericapiserver.go:342] Skipping API admissionregistration.k8s.io/v1alpha1 because it has no resources.

[restful] 2018/07/20 16:37:09 log.go:33: [restful/swagger] listing is available at

[restful] 2018/07/20 16:37:09 log.go:33: [restful/swagger] is mapped to folder /swagger-ui/

[restful] 2018/07/20 16:37:52 log.go:33: [restful/swagger] listing is available at

[restful] 2018/07/20 16:37:52 log.go:33: [restful/swagger] is mapped to folder /swagger-ui/


Worst case, we haven’t really done anything yet and you can start over with “minikube delete”, then delete the .minikube directory (likely located in %USERPROFILE%), and start over.

Once you have updated the Hyper-V configuration and started the cluster, you should be able to access the kubernetes dashboard

Actually using it

Now that you have minikube running, you can access the dashboard via a web URL – or just type “minikube dashboard” to have the site launched in your default browser.

Create a deployment – we’ll use the nginx sample image here

Voila, under Workloads => Deployments, you should see this test deployment (if the Pods column has 0/1, the image has not completely started … wait for it!)

Under Workloads=>Pods, you can select the sample. In the upper right-hand corner, there are buttons to shell into the Pod as well as view logs from the Pod.

Expose the deployment as a service. You can use the web GUI to verify the service or “kubectl describe service servicename

Either method provides the TCP port to access the service. Access the URL in a browser. Voila, a web site:

Viewing the Pod logs should now show the web server access logs.

That’s all fine and good, but there are dozens of other ways to bring up a quick web server. Using Docker directly. Magic cloudy hosting services. A server with a web server on it. K8 allows you to quickly scale the deployment – specify the number of replicas you want and you’ve got them:

Describing the service, you will see multiple endpoints.

What do I really have?

You’ve got containers – either your own container for your application or some test container. Following these instructions, we’ve got a test container that serves up a simple web page.

You’ve got a Pod – one or more containers are run in a Pod. A pod exists on a single machine, so all containers within a Pod share resources. This is good thing if the containers interact with each other (shared resources speed up this communication), but it’s a bad thing if the containers have no correlation but run high I/O functions (shared resources create contention for this communication).

You’ve got a deployment – a managed group of Pods. Each application or microservice will have a deployment. The deployment keeps the desired number of instances running – if an instance is not healthy, it is terminated and a new instance spawned. You can resize the deployment on a schedule, or you can use load metrics to manage capacity.

You’ve got services – services map resources running within pods to internal or external access. The service has an IP address and port for client access, and requests are load balanced across healthy, running Pods. In our case, we are using NodePort, and “kubectl describe service ngnix-sample” will provide the port number.

Because client access is performed through the service, you can perform “rolling updates” by setting a new image (and even roll back if the newly deployed image is malfunctioning). To roll a new image into service, use “kubectl set image deployments/ngnix-sample ngnix-sample=something/image:v5”. Using “kubectl get pods”, you can see replicas come online with the new image and ones with the old image terminate. Or, for a quick summary of the rollout status, run “kubectl rollout status deployment nginx-sample”

If the new container fails to load, or if adverse behavior is experienced, you can run “kubectl rollout undo deployment nginx-sample” to revert to the previous working container image.

When you are done with your sandbox, you can stop it using “minikube stop”, and “minikube start” will bring the sandbox back online.

A “real world” deployment would have multiple servers (physical, virtual, or a combination thereof) essentially serving as a resource pool. You wouldn’t manually scale deployments either.

Notice that the dashboard – and all of its administrative functions – are open to the world. A “real world” deployment would either include something like OpenUnison to authenticate through ADFS or some web hook that performs LDAP authentication and provides an access token.

And there’s no reason to use kubectl to manually deploy updates. Commit your changes into the git repository. Jenkins picks up the changes, runs the Maven build and tests, and creates a Docker build. The final step within the Jenkins workflow is to perform the image rollout. This means you can have a new image deployed within minutes (actual time depends on the build/test time) of committing code to a repo.

Containerized Development v/s Microservices

While both monolithic and microservice applications can be deployed in containers, there is a significant difference. Understanding that difference can save time/money/effort decomposing an application into microservices when the benefits you desire can be gained through simple containerized deployments.

One of the touted benefits of microservices — the ability for different teams to use different internal practices, different coding standards, hell even different languages and still have a functioning application because the interface is static and well documented … well, that sounds like a nightmare to me.

A company with which I worked a decade ago had teams of developers devoted to different components of the application — essentially your team owned a class or set of functions. The class/functions were had well documented and static interfaces — you wouldn’t change void functionX(int iVariable, string strOtherVariable) to return boolean values. Or to randomly add inputs (although functions were overloaded). Developers were tasked with ensuring backwards compatibility of their classes and functions. The company had a “shared libraries” development team who worked on, well, shared libraries. Database I/O stuff, authentication frameworks, GUI interfaces. A new project would immediately pull in the relevant shared functions, then start developing their code.

Developers were able to focus on a small component of the application, were able to implement code changes without having to coordinate with other teams, and consumers of their resource were able to rely on the consistent input and output of the functions as well as consistent representation of class objects.

When a specific project encountered resource shortfalls (be that family emergencies reducing workers or sales teams making overly optimistic commitments), the dozens of C# programmers could be shifted around to expand a team. In a team with an outstanding team lead, employees could easily move to other groups to progress their career.

What happens in a microservices environment? You’ve got a C# team, a Java team, a Python team. You get some guy in from Uni and he’s starting up a LISP team because Lisplets will get his code delivered through Tomcat. The next guy who comes in starts the F90 team because why not? Now I’m not saying someone with a decade of experience in Java couldn’t learn LISP … but you go back to “Google up how to do X in LISP” programming speed. There are language nuances of which you are not aware and you introduce inefficiency and possibly bugs to the code.

What’s my point? Well, (1) business practices (we program in this language, here’s our style guide, etc) are going to negate some of the perceived benefits of microservices. The small gain to be had by individual teams picking their own way are going to be outweighed by siloing (some guy from the Java team isn’t going to move into a lead role over on the C# team) and resource limitations (I cannot reallocate resources temporarily). But (2) you can architect your project to provide, basically, the same benefits.

Microservices make sense where an application has different components with different utilization rates. A product that runs a Super Bowl commercial may see a huge spike in web traffic — but scaling up thousands of complete web servers to handle the load is an inefficient use of resources. There’s a lot of product browsing, but shipping quotes, new account creations, and check-outs are not all scaling linearly to web hits. Adding tens of thousands of browsing components and only expanding the new-account-creation or checkout services as visitors decide to make purchases can be done more quickly to respond in real-time to traffic increases.

Applications where each component gets about the same amount of use … I use Kubernetes to manage a cluster of sendmail servers. As mail traffic increases, additional PODs are brought online. It’s a configuration I’d like to mirror at work — we currently have nine sendmail servers — to provide physical and site redundancy for both employee mail traffic and automated system traffic. With Kubernetes, three servers in each of the two sites (six total) would provide ample resources to accommodate mail flow. Automated systems send a lot of mail at night, and the number of pods servicing that VIP would increase. User mail flow increases during the day, so while automated mailflow pods would be spun down … user mail flow ones would be spun up. With a 33% reduction in servers, I’ve created a solution with more capacity for highly used functions (this function could be the primary usage of all six servers) that is geo-redundant (one of the current systems is *not* geo-redundant as the additional two servers in the alternate site couldn’t be justified). But I didn’t need to decompose sendmail into microservices to achieve this. Simply needed to build a containerized sendmail.