#kubernetes (2019-01)

Hello. I’m curious if anyone has had performance issues running kubectl against an EKS cluster? kubectl get po takes 5 seconds to complete. FWIW, when I used kops to create the cluster, kubectl get po would return quickly.

hrmmmmm

same size nodes, and same number of pods?

(roughly)

…are you using IAM authenticator with both?

actually, worker nodes are bigger.

let me confirm IAM authenticator

yep. uses aws-iam-authenticator.

so kops uses aws-iam-authenticator as well…

hrm…

@Andriy Knysh (Cloud Posse) have you noticed this?

(btw, are you using our terraform modules for EKS?)

sorry, no, at least not yet.

i wanted to find out if this is an EKS thing in general.

When I was testing EKS, I didn’t notice any delay

okay, that’s a good data point. thanks.

(@Andriy Knysh (Cloud Posse) wrote all of our EKS terraform modules)

https://github.com/cloudposse/terraform-aws-eks-cluster

Maybe the Authenticator is slow to connect to AWS

i’ll investigate that. thanks.

Also, how do you access the kubeconfig file?

default ~/.kube/config

something must not be configured properly. i’m investigating. i’ll let you know what i discover.

sometimes using strace helps me figure out what the process is doing

enough to dig deeper

@webb has joined the channel

Months ago, I shared a link to this awesome write up: https://medium.com/kubecost/effectively-managing-kubernetes-with-cost-monitoring-96b54464e419

I saw a demo of this yesterday and am super impressed.

I’ve invited @webb to #kubecost, so if you have any questions ping him.

Thanks for the kind words, @Erik Osterman (Cloud Posse)! We’re ready & available to help with tuning kube infrastructure!

@Erik Osterman (Cloud Posse) Check this out. New Year it the time to imagine it. https://blog.giantswarm.io/the-state-of-kubernetes-2019/

Just got access to the GKE Serverless Add-on beta: https://cloud.google.com/knative/

wow thats slick

i’m going to give it a spin… looks interesting!

feels a little bit like fargate is to ECS

yea, that’s how interpreted it

hah, we’ve all heard of dind (docker in docker)

first time hearing kind: https://github.com/kubernetes-sigs/kind

hi everyone ! I’m struggling to implement a CI/CD with Gitlab… I do have several different k8s cluster (one per stage “test”, “dev”, “stg” and “prd”) on different aws accounts (one per stage as before). I cannot find help on 2 things: how to target a specific cluster depending the branch ? and since we’re working with micro-services: how to keep a running version of my deployments on each cluster with a generic name not depending the branches names; but allowing an auto-deploy with uniques names in only one stage ? Could someone help me or link me to a good read/video about it ? right now, I just have my fresh new cluster; I still have to install/config everything (using helm).

hahaha there’s your problem.

I’m struggling to implement a CI/CD with Gitlab…

We highly recommend #codefresh.

https://codefresh.io/continuous-integration/codefresh-versus-gitlabci/

Codefresh makes it trivial to select the cluster.

We’ve used different strategies.

e.g. for release tags

1.2.3-prod or 1.2.3-staging

for branches, I suggest using a convention.

e.g.

a branch called staging/fix-widgets would go to the staging cluster

how to keep a running version of my deployments on each cluster with a generic name not depending the branches names;

Oops. Missed that.

So, the meta data needs to come from somewhere.

It can be ENVs in the pipeline configuration.

It can be branch or tag names. Note, you can use tags for non-production releases.

It can be manual when you trigger the deployments

@frednotet

cause Codefresh designed for using within Kubernetes, when Gitlab more general purpose

yes, exactly..

built from the ground up with support for docker, compose, swam, kubernetes, and helm.

Thanks I’m reading

(I just achieved my integration of gitlab but indeed I still have this multiple cluster that requires me to take the gitlab EE)

@Ajay Tripathy has joined the channel

anyone have authentication problems using metrics-server with a kops cluster? Also wondering if anyones run into heapster continuously in a CrashLoopBackOff because of OOMKilled

I’ve tried increasing the mem limit on the heapster pod but it doesn’t seem to increase

I have seen that. I recall not being able to figure it out. We don’t have it happening any more. This was also on an older 1.9 kops cluster.

it was driving me mad

no matter how much memory I gave it, it had no effect

@Erik Osterman (Cloud Posse) how’d you fix it? im on 1.11.6 kops

also have you switched to metrics-server

all our configurations are here:

https://github.com/cloudposse/helmfiles/tree/master/releases

i never ended up fixing it on that cluster. it was a throw away.

you use prom insteaD?

we do

oh wait

and heapster

you dont use heapster-nanny?

i don’t know the details

@Igor Rodionov would probably

the OOMKilled is also driving me mad

yea, sorry man!

i literally spent days on it

and didn’t figure it out

@Daren I forgot who is doing your prometheus stuff

I was having this problem on one of your clusters.

do you guys know why when I try to edit a deployment with kubectl edit the changes I make don’t stick?

usually it will emit an error when you exit kubectl edit

if it doesn’t check $?

kubectl edit ....; echo $?

weird

even if I increase the heapster deployment resource memory limit, it keeps dropping back down to 284Mi

no error btw @Erik Osterman (Cloud Posse)

$ k edit deployment heapster -n kube-system
deployment.extensions/heapster edited

@Erik Osterman (Cloud Posse) @btai we did have the heapster issue. I believe it was traced to having too many old pods for it to handle

It tried to load the state of every pod include dead ones

too many pods?

we do have alot of pods

were u able to fix it daren via configuration?

also, i think daren is talking about exited pods

Yes

@Daren since youre using kube-state-metrics, are you unable to use k top anymore

Honestly, Ive never used it, and it appears it does not work

# kubectl top pod
Error from server (NotFound): the server could not find the requested resource (get services http)

i see

https://github.com/stakater/IngressMonitorController pretty cool add-on for k8s to automatically provision health checks in 3rd party apps, these folks make a lot great open source projects, worth checking out

Yes, stakater is cool

I’ve been following them too

I want to deploy https://github.com/stakater/Forecastle

Same here, I’ve been working on a “getting started on kubernetes” blog and was looking for fun new projects to include

I’ve been trying new projects out on a Digital Ocean K8s cluster, it’s multi-master + 2 workers, 100gb storage, and a LB for $30 a month

that’s cool

not too shabby for development

@Igor Rodionov has been doing that too

It’s honestly a very nice experience, as you know, my setup at work is very smooth already

haha, that said, always want to make things smoother

I think the ease-of-use of GKE/digital ocean k8s is what we aspire to

while at the same time getting the IaC control

Yeah! It’s really nice to have the model to work off of

Especially for smaller teams that don’t need all the bells and whistles and ultimate control over every little thing

Agreed. My experience with GKE was so nice and smooth, very much so what I base a lot of our tools off of. Their cloud shell is very similar in function to Geodesic, as you’re probably aware

Yea, I saw that. I haven’t gone deep on it, but it validates the pattern.

Also, #geodesic is always positioned as a superset of other tools, which means the google cloudshell fits well inside

But you bring up a good point.

I think we can improve our messaging by comparing geodesic to the google cloud shell

yeah, at least as an introduction to the idea

@Max Moon @erik means I also use DO for my pet projects

Right! We should chat

hi everyone, i am creating a deployment example of nginx on kubernetes using manifest file and I want add prometheus monitoring on it

do you have some github manifest t share?

Are you using prometheus operator?

We use helmfile + prometheus operator to deploy monitoring for nginx-ingress here: https://github.com/cloudposse/helmfiles/blob/master/releases/nginx-ingress.yaml#L156

@Erik Osterman (Cloud Posse) yes, I am using prometheus operator pattern

@Erik Osterman (Cloud Posse) I see. I will analyse your code. I would like just to add monitoring on top a easy nginx deployment like https://raw.githubusercontent.com/kubernetes/website/master/content/en/examples/controllers/nginx-deployment.yaml

I suggest using helm for repeatable deployments rather than raw resources

(unless this is just a learning exercise)

We install the official nginx-ingress helm chart here: https://github.com/cloudposse/helmfiles/blob/master/releases/nginx-ingress.yaml

(helmfile is a declarative way of deploying helm charts)

@Erik Osterman (Cloud Posse) yes, I am using helm. I was just trying to arrange an example based on manifest

if I want to ssh into my EKS worker node, the default username is ec2user right?

@btai sorry - @Andriy Knysh (Cloud Posse) is heads down today on another project for a deadline on friday

have you made some headway?

i figured out the ssh username, i just left my question in case someone else searches for it in the future

yep! that’s great.

We’re about to release our public slack archives (hopefully EOW)

and once i got into my worker nodes, i was able to debug my issues

i do have a suggestion for https://github.com/cloudposse/terraform-root-modules/blob/master/aws/eks/eks.tf

I would change those subnet_ids to be private subnets and add a bastion module

i can help you guys do that if you’d like

yea, that’s a good suggestion.

We’d accept a PR for that (just saying…)

haha

@Erik Osterman (Cloud Posse) i remember u mentioning an aws iam auth provider that we should use for kubernetes

which one was it? kube2iam?

kiam

(avoid kube2iam)

we have an example of deploying it in our helmfiles distribution

@btai are you unblocked? what was the issue with worker nodes not able to access the cluster?

yes @Andriy Knysh (Cloud Posse) it was a stupid mistake, i had created the eks cluster security group but didn’t attach it to the cluster

@Erik Osterman (Cloud Posse) what was the reasoning to avoid kube2iam?

sec

thought we had a write up on it.

can’t find it

so kube2iam has a very primitive model. every node runs an a daemon.

when a pod needs an IAM session, it queries the metadata api which is intercepted by iptables rules and routed to the kube2iam daemon

that part is fine. that’s how kiam works more or less.

the problem is if you run a lot of pods, kube2iam will DoS AWS

AWS doesn’t like that and blocks you. so the pod gets rescheduled to another node (or re-re-scheduled) until starts

so we have this cascading problem, where one-by-one each node starts triggering rate limits

and then it doesn’t back off

so now we have 5000 pods request IAM credential in an an aggresive manner and basically the whole AWS account is hosed.

kiam has a client / server model

you run the servers on the masters. they are the only ones that need IAM permissions.

the clients request a session from the servers. the servers cache those sessions.

this reduces the number of instances hitting the AWS IAM APIs

and results in (a) faster assumed rules (b) less risk of tripping rate limits

https://docs.cloudposse.com/troubleshooting/kube2iam-assuming-role-access-denied/

@Erik Osterman (Cloud Posse) awesome that makes sense. thanks for the detailed answer!

hi everyone, I am deploying a prometheus operator to monitor my application. I probably misunderstood how it works

basically, for each application or servicemonitor you will have a prometheus instance

or you can share the cluster one with your application? what is the practice?

@deftunix when you deploy prometheus operator, it will scrape all pods including the app, so you don’t need to anything special about it

here’s how we deploy it with helmfile https://github.com/cloudposse/helmfiles/blob/master/releases/prometheus-operator.yaml

it will create these resources

https://github.com/helm/charts/tree/master/stable/prometheus-operator/templates

yes, I have the prometheus operator running and monitoring my base infrastructure

I deployed it using the coreos helm chart in a monitoring namespace but my application service are not scraped

it’s scarpping just a set of servicemonitors “seems” predefined

does the app output any logs into stdout?

yes! I deployed an nginx with the exporter. when I created with the operator a servicemonitor and prometheus instance

dedicated to the app, it works

the target appear

https://github.com/coreos/prometheus-operator/blob/master/Documentation/user-guides/getting-started.md

I am following this

but I was expecting that adding the annotation to the services the scrape was automatic

and new target will be showed in my target list

did you also deploy kube-prometheus?

https://github.com/coreos/prometheus-operator#prometheus-operator-vs-kube-prometheus

https://github.com/cloudposse/helmfiles/blob/master/releases/kube-prometheus.yaml

in my “cluster-metrics” prometheus yes

so when you install kube-prometheus, it will install a bunch of resources including https://github.com/prometheus/node_exporter

which will scrape metrics

yes, from node, apiserver, kubelets, kube-statistics

my problem are not the cluster-metrics, because them are fully supported by default by the helm chart but understand how the operator pattern work

TIL: https://github.com/kubernetes/kubernetes/issues/57291

Would have assumed the threshold off a CrashLoopBackoff be configurable

I am working on a demo where we deliberably kill pods

so I want to show resiliency. oh well.

have you guys checked out https://github.com/windmilleng/tilt

https://github.com/kubernetes-sigs/kubebuilder

https://github.com/iJanki/kubecron

@Daren