Episode

Kubernetes at LinkedIn, with Ahmet Alp Balkan and Ronak Nathani

Podcast: Kubernetes Podcast from Google
Published: Mar 25, 2025
Duration seconds: 2522
Processing state: processed
Canonical source: https://e780d51f-f115-44a6-8252-aed9216bb521.libsyn.com/kubernetes-at-linkedin-with-ahmet-alp-balkan-and-ronak-nathani
Audio: https://traffic.libsyn.com/secure/e780d51f-f115-44a6-8252-aed9216bb521/KPod249.mp3?dest-id=3486674
JSON: /v1/public/podcasts/kubernetes-podcast-from-google/episodes/kubernetes-at-linkedin-with-ahmet-alp-balkan-and-ronak-nathani
Markdown: /podcast/kubernetes-podcast-from-google/kubernetes-at-linkedin-with-ahmet-alp-balkan-and-ronak-nathani.md

Actions

POST https://stenobird.com/v1/public/podcasts/kubernetes-podcast-from-google/episodes/kubernetes-at-linkedin-with-ahmet-alp-balkan-and-ronak-nathani/transcription-requests
Idempotently request low-priority transcript generation for this episode.
GET https://stenobird.com/podcast/kubernetes-podcast-from-google/kubernetes-at-linkedin-with-ahmet-alp-balkan-and-ronak-nathani.md
Read the agent-friendly Markdown representation of this episode resource.

Summary

LinkedIn engineers reveal how they manage massive Kubernetes clusters on bare metal, focusing on handling latency-sensitive workloads and hardware fragmentation. The discussion explores the complexities of building an opinionated platform that abstracts infrastructure for developers while maintaining high performance.

Topics

Kubernetes
Bare Metal
Infrastructure at Scale
Cloud Native
Custom Controllers
LinkedIn Engineering
Site Reliability Engineering
Container Orchestration

Highlights

Main idea: LinkedIn uses a custom, opinionated platform on bare metal to manage hardware generations and prevent fragmentation
Practical takeaway: Use node profiles and specific pools to isolate latency-sensitive applications from older hardware generations
Failure mode: Large-scale clusters can cause controller informers to fail to sync, potentially leading to catastrophic state loss like label clearing
Main idea: Kubernetes can be adapted for stateful workloads if you control the full stack from bare metal to disk attachment
Practical takeaway: Abstracting Kubernetes complexity via custom APIs helps developers focus on resources like CPU and memory rather than low-level networking

Chapters

1:00 Kubernetes Ecosystem News: Updates on CubeFS graduation to CNCF and Canonical's announcement regarding 12-year Kubernetes Long Term Support.
4:10 Running Kubernetes on Bare Metal: An exploration of LinkedIn's transition from legacy containerization to a large-scale Kubernetes implementation on bare metal.
10:25 Networking and Flat Data Center Architectures: Discussion on bypassing standard CNI plugins like Flannel in favor of a flat, routable data center network.
16:40 Managing Hardware Skews and Fragmentation: How to handle different CPU generations and the challenges of preventing resource fragmentation in multi-tenant pools.
22:50 The Power of Custom Controllers: The necessity and reality of developing custom Kubernetes controllers and CRDs to extend platform functionality.
32:15 Abstracting Complexity for Developers: Providing an opinionated interface that allows engineers to specify compute resources without managing low-level Kubernetes primitives.
38:45 Lessons from Large-Scale Failures: A deep dive into how controller informer timeouts in massive clusters can lead to unintended side effects like mass label deletion.