Tim Ebert

Cloud Engineer @stackitcloud working on #Kubernetes @GardenerProject 👨🏻‍💻👨🏻‍🌾🌱 | Ice cream and board game fanatic🍦🃏🎲

[Tim Ebert @timebertt]

@alexellisuk Not in production, but the evaluation in my thesis showed that it scales well. You can check out the results in the docs or the linked thesis repository

[Tim Ebert @timebertt]

I’m at #KubeCon this week! Come by the STACKIT booth and say hi. I‘m happy to chat about sharding in Kubernetes controllers: https://t.co/VeDNse9wBG

[Tim Ebert @timebertt]

I recently submitted my Master’s thesis on sharding Kubernetes controllers for making them horizontally scalable. Read the full thesis here: https://t.co/c7IRcBfEJO

[Tim Ebert @timebertt]

Super happy about how my Master's project on #Kubernetes controller sharding turned out: https://t.co/VeDNse9wBG
It's ready for usage and experimentation. Check it out, give it a spin, and reach out with feedback!

[Tim Ebert @timebertt]

@the_sttts @ibuildthecloud @Mangirdas What would be the limit then? etcd or API server? I guess the controller-side can scale out nicely as long as the watch cache is restricted to what the controller shard is actually responsible for.

[Tim Ebert @timebertt]

@the_sttts @ibuildthecloud @Mangirdas Maybe sharding controllers and their caches can help? Some ideas I'm experimenting with these days: https://t.co/VeDNse9wBG

[Tim Ebert @timebertt]

@alexellisuk @tsaha Leader election prevents you from scaling controllers horizontally. I looked into the topic for a study project and working on it again right now for my Master’s thesis: https://t.co/VeDNse9wBG
The idea is to apply sharding mechanisms to make controllers scalable

[Tim Ebert @timebertt]

@ibuildthecloud One can instruct API server to write leases to a different etcd cluster and configure this one to use the in-memory store. We are using this option to write events to a dedicated etcd cluster which still has persistent storage but doesn’t do costly delta snapshots.

[Tim Ebert @timebertt]

@alexellisuk @honratha Yes I think so

[Tim Ebert @timebertt]

Ready for #KubeCon #Amsterdam, see you tomorrow!

[Tim Ebert @timebertt]

Talking at #Rejekts23 was fun! 🎉 🙌🏻 Let me do another round! 😅
I opted in for a spontaneous lightning talk about my study project on Kubernetes Controller Sharding at 6:15 pm.
If you’re interested, here is the live stream: https://t.co/0jaLR8K9vA

[Tim Ebert @timebertt]

I’m excited to be speaking at @rejektsio in Amsterdam! Join our talk in which we will share our learnings from implementing credentials rotation for #Kubernetes clusters at scale in @GardenerProject. Looking forward to having a chat at the event!

[Cloud Native Rejekts @rejektsio]

We welcome @rafaelfranzke (@SAPdach) & @timebertt (Stackit) to #Rejekts2023 in Amsterdam!🌷🚴

They will walk us through methods to manage a plethora of credentials in a secure way & share learnings from cluster admins as well as app devs!

For late🎫s👉https://t.co/LlqJAF5785

[Tim Ebert @timebertt]

@stefanprodan @fluxcd I explored the idea of controller sharding last year in a study project: https://t.co/VeDNse9wBG
It builds on the same idea (label selector per instance) but doesn't require explicit configuration and also adds an instance that takes over labeling. Might be interesting for you :)

[Tim Ebert @timebertt]

@thockin @alexellisuk @crenshawdotdev @CloudNativeFdn Yep, then we are back at trimming off the suffix :D

[Tim Ebert @timebertt]

@thockin @alexellisuk @crenshawdotdev @CloudNativeFdn Right, if you know how many levels you want to get up you could do as many joins

[Tim Ebert @timebertt]

@thockin @alexellisuk @crenshawdotdev @CloudNativeFdn If you had metrics for owner references you were able to join time series in PromQL for grouping. Owner references as such are core api machinery concepts and applicable to all resources. Generically finding the top-most owner is probably not possible in PromQL

[Tim Ebert @timebertt]

@alexellisuk @honratha There is a containerd cri configuration option which tells it to pull images for registry x from mirror y: https://t.co/UxNuDxuDN0
We didn’t try supporting it for the legacy dockershim, so I don’t know if this works with docker as well.

[Tim Ebert @timebertt]

@alexellisuk @honratha We ended up deploying a cache per upstream registry listening on different ports as well. So far this works great for speeding up image pulls for different public upstream registries in our dev setup.

[Tim Ebert @timebertt]

@alexellisuk @honratha I think this might be a limitation in the protocol. I remember looking into the same thing and checking debug logs how things are actually fetched from the mirror: the host (registry name+port) of the image name was never sent over the wire but only the repository+tag.

[ahmet alp balkan @ahmetb]

Seattle Kubernetes Meetup is back after 3 years!🥳 https://t.co/K6MQN5BuPe I'll be there to talk about scaling controllers and present @timebertt's paper on this topic.