EFK (Elasticsearch-Fluentd-Kibana) in kubeadm-dind-cluster

When you wish to have your logs sent to Elasticsearch and browse them with Kibana you should add to your k8s cluster EFK stack.

My first attempt to install EFK was with https://github.com/kubernetes/kubernetes/tree/master/cluster/addons/fluentd-elasticsearch

However it was fluentd was failing to send logs to Elasticsearch.

Fortunately on digitalocean website there is a blog post describing on how to do it and it happens to be working on a dind cluster. https://www.digitalocean.com/community/tutorials/how-to-set-up-an-elasticsearch-fluentd-and-kibana-efk-logging-stack-on-kubernetes

All the config posted over there is valid except for few caveats related to how dind stores logs.  To make fluentd pods have access to logs in fluentd.yaml add in volumeMounts:

And in volumes:

Also in elasticsearch-statefulset.yaml decrease required storage space and match storageClassName to the one that exists in the cluster in volumeClaimTemplates:

Here local-path provisioner has been used and storage is set to a low value like 1Gi.

Local-path provisioner has been added to the cluster (https://github.com/rancher/local-path-provisioner)

Now EFK can be onboarded onto a cluster as follows:

Check if you can access elasticsearch and kibana:

After port-forward provide container name relevant to your cluster. Check access on a host running kubectl port-forward

Now your system logs (not related to user pods, like kube-apiserver) should be fed by fluentd to elasticsearch and kibana can read them for es backend. You can have workload pods logs logs flowing into elasticsearch or build a dedicated instance of EFK just for that purpose to segregate system vs workload logging.

 

kubeadm-dind-cluster or kind

Even though it was recently retired in favour of kind (https://kind.sigs.k8s.io/kubeadm-dind (https://github.com/kubernetes-retired/kubeadm-dind-cluster) is still a great way to run a kubernetes cluster locally on laptop (in my case it is MacBook). It’s main limitation is that it supports v1.14.1, while kind supports v1.15.3 (though yesterday v1.16 version got released). Both rely on the same principle that nodes are not baremetal nodes nor VM, they are plain docker containers.  Then kubeadm is used to deploy k8s clusters on docker-based k8s cluster. System pods as well as user pods run as embedded containers inside docker container hosts (hence dind = docker-in-docker)

Yet with kubeadm-dind-cluster you can easily switch between CNI plugins, and it works between Docker for Desktop restarts and OS reboots.

Switching between CNI plugins is as easy as modifying inside an install script the following variable:

Main issue with kind is that OS reboot (not necessarily docker restart) causes originally assigned IPs to docker containers acting as k8s nodes to be changed. Hence after creating a cluster, making some configuration changes, or deploying pods etc we are forced to redo the job. Root cause of the problem is that docker doesn’t maintain originally assigned IPs to containers across docker restarts/OS reboots – at least it is not forced to do it (https://github.com/moby/moby/issues/2801). It is possible to assign static IPs to containers when starting containers with docker run. Unfortunately kind does not offer such option (for now). Issue is tracked at: https://github.com/kubernetes-sigs/kind/issues/148

All in all, when you want to practise with k8s (multi-node) on your laptop kubeadm-dind-cluster. kind is a very near future when it is made stable across docker restarts. 

Terraform or Pulumi for IaC

Terraform by Hashicorp is a pioneer in the area of IaC tooling which allows to define infrastructure in a declarative manner by using its own language (HCL – Hashicorp Configuration Language). That means:

  • it takes definition of the infrastructure (simplistic example could be a couple of VMs running in AWS or OpenStack) found in the config file (can be a single .tf file) defined in HCL
  • connects to the IaaS by using an appropriate provider – plugin/extension that allows to talk to AWS/Openstack etc.
  • compares current state of the infrastructure against what is found in the definition file (.tf)
  • tries to meet the desired state found definition file (using example below it means to deploy VMs)

Approach as above allows to have a deterministic state of your infrastructure and deployment is performed by calling API of the infrastructure provider rather than manually or using different methods already offered by them (e.g. HEAT templates for Openstack, python client vs CloudFormation for AWS, awscli tool). If a VM is deleted for any reason on the IaaS re-running the terraform apply (command to check current state of the infra, comparing against the definition file and trying to meet the desired state from the file) will just deploy the missing VM. 

It allows to have a common language for defining the infrastructure on different infrastructure providers (bearing in mind that each of them use basic different building blocks to create the infrastructure e.g. aws_subnet resource vs openstack_networking_subnet_v2 resource). 

Terraform allows for more than just creating resources on public/private cloud providers, it can deploy on kubernetes (more of a PaaS than IaaS), or interact with RESTAPI. People/companies create their custom providers.

However, anyone with some development background will see that that HCL language is limited – even writing a simple loop, while possible – it is somehow done indirectly (I admit I did not have time to check 0.12 option which was supposed to add some typical programming constructs like for loop). Immediately the feeling is that it should be done differently – you want to write a code and have appropriate programming libraries to wrap API calls to different providers. This is where Pulumi comes in – its tenets are the same:

  • it is declarative
  • it compares the current state against the definition (defined in the programming language of choice – at the moment of writing JS/TS/Python/Golang are supported)
  • uses programming language libraries to connect to IaaS (e.g. pulumi_gcp in Python) which extend core pulumi framework
  • tries to meet the desired state (pulumi up vs terraform apply)

What is new is that it’s SaaS by default and requires to setup an account on Pulumi website. Pulumi service keeps state on Amazon S3, but there is also an option to keep it locally (different backends). If it is kept on Pulumi then you can check your created infrastructure resources via their website (actually this is very cool, you can even click to connect to created VMs consoles’).

Rather than having a separate language to define the infrastructure you just import appropriate libraries and start defining the infrastructure directly in code. There is no better source of introduction to Pulumi than Joe Duffy’s blog post: http://joeduffyblog.com/2018/06/18/hello-pulumi/.

So which one should you choose? I would say it depends on the people who are going to use it:

  • Developers, people with development skills should opt for Pulumi so as not to be constrained by HCL. That also means learning curve is steeper for Pulumi.
  • People without such skills will feel more comfortable with HCL and might not feel constrained by it. They should go with Terraform. I just felt I had to create a lot of boilerplate with it.
  • Terraform is more mature (older tool) but Pulumi’s people are quite helpful and available on Slack so it is just kicking off.

I have posted few examples with both tools in my gitlab repo:

https://gitlab.com/stackblog/terraform

https://gitlab.com/stackblog/pulumi

Juniper vMX – MPLS over lt interface

I was wondering if I can run MPLS over OVS, VXLAN tenant networks in Openstack (for example VLAN tagging won’t work in such case, not by default at least, didn’t have an option to test with vlan_transparent setting on network level). Wanted to quickly check it out between a pair of vMXes with static LSPs remembering I can run logical-systems or routing-instances interconnected between themselves with tunnel interface (lt). That works just fine but only with logical-systems.

LS-vmx01 lt-0/0/0.1<–>lt-0/0/0.2 vmx01 ge-0/0/2.0<–>ge-0/0/2.0 vmx02 lt-0/0/0.2<—lt-0/0/0.1 LS-vmx02

vmx01 cfg:

vmx02 cfg:

Main thing here was to set family mpls on lt-0/0/0 units and add lt-0/0/0.2 interface to protocols mpls.

Only after that POP entries appeared in routing and forwarding tables:

vSphere ESXi access without vSphere client

Today I stumbled on quite irritating problem. If you are OSX or Linux user and your ESXi setup doesn’t have vCenter (which comes with a price) then you are:

  1. Doomed to run a VM with Windows and install a vSphere client there (it requires Windows and .NET)
  2. Lose your time over trying to run it using Wine (there are instances where where with specific wine version and vSphere client it could work – I didn’t manage) or Crossover (if on Mac)
  3. Install VMware Fusion (if on Mac) – but it is a paid solution, on the other hand you might already have it – it is capable of connecting to vCenter and ESXi hosts.
  4. Install ESXI host client UI https://labs.vmware.com/flings/esxi-embedded-host-client

Following assumes there is internet connectivity from your ESXi host (but that’s not a big deal if there is not, just SCP to upload vib package), so connect to your ESXi host over SSH and then:

Now you can access your ESXi host over UI: https://<esxi_host_ip/ui/

ConfD – making your network elements programmable

Tail-F, originally a Swedish company bought over by Cisco Systems, created a management agent software framework for network elements and applications – ConfD. If employed as a part of software – ConfD allows to be programmability an inherent part of it. This post will us a very tangible example of its power – while Linux natively doesn’t offer NETCONF to manage interfaces – with use of ConfD and special program that makes use of it – programmability is made possible.

ConfD uses, what most of up-to-speed network engineers should know, YANG – a data modeling language for NETCONF (if we treat NETCONF as a new SNMP then YANG is like SMI – “special version” of ASN.1 was for SNMP) – better description can be found here: http://www.tail-f.com/what-is-yang/

The way to depict what ConfD does is – let’s say if you create your own program that acts as a router and you need a CLI for e.g. show ip route and additionally you would like to have a possibility to check routes over the NETCONF then it’s going to be done automatically for you. ConfD will run as a process next to your routing process and will manage it’s configuration, offer CLI and NETCONF (also RESTCONF/SNMP/WebAPI if Premium version is used). Of course that means that a if you want to implement ConfD shall be a part of implementing your software, whole configuration, CLIs etc.

There are 2 versions of confd available – free (basic) and more capable (paid) premium – comparison can be found here: http://www.tail-f.com/confd-basic

More on how ConfD works and its composition (for example there is a database – CDB – that keeps the config) can be found on Tail-F site.

Let’s get to business, demonstration of how it can be used is based on:

  • confd basic 6.3
  • ydk 0.7.1 (there will be a separate post on ydk, installation, usage and what is it)

The very reason for using confd 6.3 is that newer versions i.e. 6.4 and 6.6 use YDK 1.1 for modelling and example program I wanted to test with YDK on newer confd versions uses YANG 1.1 while YDK has just partial support for it (full support is for YDK 1.0 – RFC 6020).

In order to install confd:

  1. Download the 6.3 and 6.4 versions appropriate for you os from https://developer.cisco.com/site/confD/downloads/
  2. Install it: 

  3. Source confdrc file (so as to have access to confd cli tools:

  4. Run some example intro example:

  5. access confd CLI by running confd_cli command unless you started using make cli that takes you directly to CLI

 

Run the example:

Example I would like to use is called linuxcfg but id doesn’t work out of the box. When both versions 6.3.and 6.4 are installed the Makefile in ipmibs directory must be overwritten in 6.3 examples folder:

When this is done you can compile the example.

ip a from linux level:

show ip ipAddressEntry from confd level:

Ok so maybe output is not ideal like… IP addressing in HEX.

You can also use pipes and regex to filter what you only need, you can save output, export to csv, json, curly-braces (like Juniper CLI) etc.

From now on you are on your own….just kiddn’ in the next post I will show how to setup YDK and query linucfg over NETCONF with YANG models provided with this example. Now we have a running linuxcfg with confd that exposes NETCONF. More to follow…

Modifying virtual machine images

What to do when you have a virtual machine image and for a example you need to some files contents like ssh config or so? Modified images can be uploaded to glance – repeating same step after running several VMs of the same type can be easily avoided in this way.

There are few tools that can be used for that purpose and are extremely powerful (most importantly, these are usually run in place :

Example – install wireshark & nmap in an RHEL minimal install image (we assume there is Internet access from a machine running virt-customize command) – below is basically running locally a VM, modifying it and saving the changes to the image.

Example – modify ssh server config. Following edits a file in a filesystems and saves the changes to the image.

Guestfish gives access to the filesystem – it is more powerful the virt-edit in a sense that it allows browsing through the filesystems rather than modifying a file that you know a path for. You can also create new files and add contents to them. Example sequence of step to perform is as follows:

1) OPTIONAL: export LIBGUESTFS_BACKEND=direct

2) guestfish -a <qcow2/img>

3) run

4) list-filesystems

5) mount <root_filesystem> /

6) modify files

7) umount /

8) exit

Increase LV size of a qcow2 image

When you get a qcow2 image with a given size – it can’t be simply changed on-the-fly while running a VM or by giving just more space to a VM flavor in OpenStack. Situation gets even more complex when image has LVs inside but fortunately by using guestfish and virt-resize image can be suited to one’s needs. Below are the steps that I used to perform such modifications:

Default image: image-name-250G.qcow2
Resized: image-name-750G.qcow2

1. Check which device to resize (this image has LVM created PV on /dev/sda2):

2. Resize image (from 250 to 750GB) – resizing is NOT performed in place:

3. Resize disk and specific device (in this case it is /dev/sda2) and LVM PV:

4. Go to guestfish and use free space on VGS to create additional LV, create additional filesystem and mount point (DISABLE 64bit flag on EXT4, required by this image as it uses outdated e2fsck
that doesn’t support 64bit option):

5. Upload image to glance.

Automatically provisioning baremetal/VM server

During my endavours I had a situation where I had to provision 10 servers (install all of them manually and configure same things on all of them, same files etc).

There is a nice alternative to it called Stacki from StackIQ (bought by Teradata last year). What it offers is specialized PXE server that is used to boot baremetal/VM servers (CentOS/Redhat/Ubuntu). 

Its architecture look as follows (Stacki server == Frontend, server to be provisioned == backend):

Firstly in CSV file you prepare a list of hosts with their MACs and as a next step you add puppet that will be used to provision the servers after booting.

More can be found here: 

https://github.com/Teradata/stacki

Frontend machine can be a VM – actually it worked pretty nice – tested with provisioning other VMs.