Vertical Pod Autoscaling

[jbeda]

https://github.com/kubernetes/autoscaler/tree/master/vertical-pod-autoscaler

In general how to set requests/limits and what that might look like going forward.

[jbeda]

Monday or Tuesday:

• [x] Pick a topic and create/use an issue in the repo
• [ ] If applicable coordinate with an upstream project if they're interested in attending the show
• [x] Create the graphic
• [x] Schedule on Youtube
• [x] Add scheduled episode to TGIK playlist on YT
• [x] Create short link in rebrandly
• [x] Tweet about it from your personal account
  • [x] Post tweets on internal #tgik slack channel and cc @castrojo and Vykky
  • [ ] Get confirmation from boschem that she's seen the tweet and is scheduling it

Before episode:

• [x] Create HackMD page for live notes. Start with README template in this repo
• [x] Paste HackMD URL on internal #tgik channel and cc @castrojo or other assistant
  • [ ] Assistant adds ~5-10 news items from around the web as possible "News of the Week"
• [x] Get everything tested well before episode so you can debug
• [x] Use internal slack channel to ensure that AV setup is good

After show (ideally right away but can wait until Monday):

• Create/submit PR for this repo
  • [x] Episode directory and README from hackmd
  • [x] Any other files from the episode that might be useful
  • [x] Update playlist.md with episode
  • Feel free to self merge
• [x] Edit YT description to point to episode directory here
• [ ] OBS/OS Upgrade if applicable (Never on the day of a show)

[jbeda]

https://studio.youtube.com/video/p6zCzyBNhLE/livestreaming

[jbeda]

https://github.com/FairwindsOps/goldilocks is relevant.

[jbeda]

Also Google paper on autopilot: https://dl.acm.org/doi/pdf/10.1145/3342195.3387524

[jbeda]

From Duffie: we need to add these two lines to the components.yaml from metrics-server to support kind https://gist.github.com/mauilion/cd10204e924600c15c5c14c97c2ca66b#file-components-yaml-L91-L92 cause the hostnames of the nodes aren't resolvable from the metrics-server pod. and because the serving certs for the kubelet are self signed. (edited)

From Jorge: Office hours video talking about VPA -- https://www.youtube.com/watch?v=5kJ6tJXq-qU&feature=youtu.be&t=2558

[jbeda]

Also from Duffie: https://medium.com/infrastructure-adventures/vertical-pod-autoscaler-deep-dive-limitations-and-real-world-examples-9195f8422724

And folks often use the kube-prometheus and the prom adapter to populate the metrics api in kube https://github.com/coreos/kube-prometheus/tree/master/manifests

[jbeda]

From @embano1: https://twitter.com/embano1/status/1257779994664861696. Some analysis on Autopilot and pointer to Cindy's take.