Or you can signify the context using shell prompt and use kube rbac to give read only permissions instead of letting people mess with your prod cluster.
Managing shell tooling to edit ps1 for this is a bad idea, because there are already many other much longer established mutations people use that for (like git branch or activated Python environment), kubernetes can’t come along decades later and suggest to use the tool effectively one has to break those long standing workflows.
It’s always an available option for people who want it, but many many use cases are poorly served by it.
The RBAC thing is different. I am 100% for those RBAC limitation as long as I have total authority to make a devops admin do what I need them to do right when I need them to do it.
If my team is going to get alerted on build failures or service errors that require kubernetes mutation actions or exec to debug or resolve (99% of the time they do), then my team needs full admin access.
If rbac prevents access, then you better route the alerts to someone else or give us authority to compel instant triage responses.
In the 3 different companies I’ve worked in that use large kubernetes clusters, rbac has been a miserable problem across the board, because SREs don’t want to be responsible for triaging or resolving application team issues, and application teams will not honor alerts that are not actionable because of poorly conceived permissions issues where people mistakenly think they are setting useful access control policy but really they aren’t.
You could have a mechanism to grant yourself temporary write access on an as-needed basis.
At Zalando everyone has read access (except to secrets). You can request write access with a command line tool another employee can then approve it. There is also an option to request access with an incident ticket, in that case you immediately get write access without approval by someone else. Write access expires after 1h.
Access to the underlying AWS account uses the same mechanism.
But then when you’ve unlocked access, you could just make the same fat-fingered mistakes due to implicit use-context settings.
You might use use-context to set yourself to that restricted production context at the start of an incident. 30 minutes later you resolved the incident, but the access is still active and whoops you delete a deployment you thought you were deleting from a stage experiment or something.
All that does is add an extra hoop to jump through. People need to stop believing that adding extra bureaucratic hoops offers any type of safety - it doesn’t.
If you want to restrict access then you need to actually restrict it and convert on-call alert responsibilities to a central devops team that, like it or not, is responsible for solving everyone else’s problems.
If you don’t want a central devops team that can be paged and on the hook for everyone else’s systems, then you cannot have write access controls.
There’s just no way out of this dilemma. Temporary access that can be self-granted == no access control. Temporary access that requires admin approval == admin team is on pager duty for every other team.
kube-ps1, kubectx and kubens can help quite a bit here. kube-ps1 to give you a visual indication of current context and the others to make switching easier.
You can also explicitly add namespaces to manifests, which help avoid applying to the wrong namespace.
While kube-ps1 helps, remember that the context is set session-wide. So even if your last command says "development" in one terminal, another terminal may have modified the context by now.
So make sure to spam the return key before deleting stuff :D
Agreed. With autocomplete I find myself being able to make those long commands painless while making sure I'm being explicit. Only took me one time running the wrong command against the wrong cluster to drill that one into my brain forever haha
I've always been confused with kubectx. How do I actually add a seperate yaml file (say the one I download with my browser from DigitalOcean) into the context?
kubea(){
if [ "$#" = 1 ]; then
KUBECONFIG=~/.kube/configs/"$1"
else
ls -1 ~/.kube/configs/
fi
}
That way I can "activate" a specific config the same way I activate my Python virtualenv, for a specific terminal. I also add the basename of the config to my prompt.
It's funny, I used to have something exactly like this, but found that too many other tools and scripts also needed to access the right cluster. Persistent state works better for me in practice.
It's a mixed bag. On one hand it's a pain when you have to switch back and forth. On the other hand it's awesome that all the state is in one, human-readable file and can easily be duplicated or pointed somewhere else, etc.
I primarily use OpenShift which fixes part of it, and with a few bash aliases the rest are pretty painless as well.
What I have been doing is a directory per namespace/context.
I have a common Makefile that I link there which creates the namespace and does all the basic initialization stuff like inserting deny-all netpol rules, credentials for our docker registry.
It also runs a puppet task that creates scripts to use kubectl and helm with the correct context.
An optional config file let's me change the defaults, like pointing it to a different cluster.
I also keep context specific notes and files, scripts, etc. in this directory. Everything under version control, except for the autogenerated stuff.
For normal usage I always just run ./kubectl.sh. I switch between contexts by changing directories and the commands to get pods and talk with the cluster don't change. I just go nuts with ctrl+r.
I also can run commands on different contexts easily, lets me pipe the output from one context into another just running:
namespace1/kubectl.sh exec ... | namespace2/kubectl2.sh run ...
namespace1/mysqldump.sh mydb | namespace2/mysql.sh myotherdb
Each kubectl.sh points to the correct version of kubectl for each cluster, so you don't get weird kubectl errors from mixing stable clusters with more modern ones.
The config file being a puppet script is a very powerful tool. I use it to for example generate the config files for our software and the yaml files for helm. Also for doing the vendor specific stuff like allocating volumes, ips and configuring DNS servers.
Super annoying why ever have very specific way to merge multiple contexts. Google utils writes to this file one way, other tools another way. So weird and so not-transparent.
I like to have one config file per cluster and point kubeconfig to all these files. Then I use a bash function to select the cluster by name. My zsh shell previews the cluster name on the right when I start typing k, kubectl, helm and similar... So it is harder for me to make a mistake
28 comments
[ 1.5 ms ] story [ 84.3 ms ] threadYou should absolutely always be typing out huge, verbose commands like kubectl —context my-context -n my-namespace <some commands> ...
The opportunity for tragic error when you have implicit context or namespace settings applied is just too great.
It’s always an available option for people who want it, but many many use cases are poorly served by it.
The RBAC thing is different. I am 100% for those RBAC limitation as long as I have total authority to make a devops admin do what I need them to do right when I need them to do it.
If my team is going to get alerted on build failures or service errors that require kubernetes mutation actions or exec to debug or resolve (99% of the time they do), then my team needs full admin access.
If rbac prevents access, then you better route the alerts to someone else or give us authority to compel instant triage responses.
In the 3 different companies I’ve worked in that use large kubernetes clusters, rbac has been a miserable problem across the board, because SREs don’t want to be responsible for triaging or resolving application team issues, and application teams will not honor alerts that are not actionable because of poorly conceived permissions issues where people mistakenly think they are setting useful access control policy but really they aren’t.
At Zalando everyone has read access (except to secrets). You can request write access with a command line tool another employee can then approve it. There is also an option to request access with an incident ticket, in that case you immediately get write access without approval by someone else. Write access expires after 1h.
Access to the underlying AWS account uses the same mechanism.
You might use use-context to set yourself to that restricted production context at the start of an incident. 30 minutes later you resolved the incident, but the access is still active and whoops you delete a deployment you thought you were deleting from a stage experiment or something.
All that does is add an extra hoop to jump through. People need to stop believing that adding extra bureaucratic hoops offers any type of safety - it doesn’t.
If you want to restrict access then you need to actually restrict it and convert on-call alert responsibilities to a central devops team that, like it or not, is responsible for solving everyone else’s problems.
If you don’t want a central devops team that can be paged and on the hook for everyone else’s systems, then you cannot have write access controls.
There’s just no way out of this dilemma. Temporary access that can be self-granted == no access control. Temporary access that requires admin approval == admin team is on pager duty for every other team.
You can also explicitly add namespaces to manifests, which help avoid applying to the wrong namespace.
So make sure to spam the return key before deleting stuff :D
I primarily use OpenShift which fixes part of it, and with a few bash aliases the rest are pretty painless as well.
I have a common Makefile that I link there which creates the namespace and does all the basic initialization stuff like inserting deny-all netpol rules, credentials for our docker registry.
It also runs a puppet task that creates scripts to use kubectl and helm with the correct context.
An optional config file let's me change the defaults, like pointing it to a different cluster.
I also keep context specific notes and files, scripts, etc. in this directory. Everything under version control, except for the autogenerated stuff.
For normal usage I always just run ./kubectl.sh. I switch between contexts by changing directories and the commands to get pods and talk with the cluster don't change. I just go nuts with ctrl+r.
I also can run commands on different contexts easily, lets me pipe the output from one context into another just running: namespace1/kubectl.sh exec ... | namespace2/kubectl2.sh run ... namespace1/mysqldump.sh mydb | namespace2/mysql.sh myotherdb
Each kubectl.sh points to the correct version of kubectl for each cluster, so you don't get weird kubectl errors from mixing stable clusters with more modern ones.
The config file being a puppet script is a very powerful tool. I use it to for example generate the config files for our software and the yaml files for helm. Also for doing the vendor specific stuff like allocating volumes, ips and configuring DNS servers.
KUBE_APISERVER_ADDR=https://foo.com KUBE_CERT_FILE=/tmp/cert KUBE_NAMESPACE=bar kubectl get po
This way, I can export them into my environment with my .bash_profile instead of having to persist state just to perform cluster ops
[1] https://kubernetes.io/docs/tasks/access-application-cluster/...