Implementing a Consumer Focused SLA for a Kubernetes Based PaaS
Talkat KubeCon + CloudNativeCon North America 2019 · on Wednesday, 2019-11-20· at 10:55· Room 11AB - San Diego Convention Center, Upper Level·
summary
Box runs ~150 services on ~25,000 pods (9,000 in the largest cluster) and moves from crude API 500 checks and synthetic probes to a consumer-facing SLA based on critical replica availability: service owners declare the minimum replicas to serve peak traffic, the platform auto-adds buffers (hardware, PDB/rolling update, maintenance) and enforces guaranteed QoS, liveness/readiness and PDBs via a kube-applier pipeline driven by JSON config. Goal is to raise effective platform availability from ~99.4% toward ~99.99%.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 0:52Overview
- 1:12History of Box
- 2:39Platform
- 3:54Crosscutting Concerns
- 4:44The Problem
- 6:57The First Solution
- 10:38Protections
- 13:04CRI
- 14:59Example
- 17:31Measuring SLA
- 19:13Future work
- 24:53Questions
- 25:08Sock compliance
- 28:16Extending to other services
- 29:55What if your egress is down
- 31:53Is it tricky for teams to define their pod count
related talks
36mContinuous Delivery with Kubernetes at Box [I]
25mAn SLO-Driven Approach to Enhance Kubernetes Cluster Reliability
46mHow Box Runs Containers in Production with Kubernetes
39mKubernetes Scalability Definition Evolution
23mScaling Kubernetes in Regulated Environments Through Platform APIs
36mOptimizing Resource Usage in Kubernetes
30mOperating within Normal Parameters: Monitoring Kubernetes
33mManaged Kubernetes Service: Day Zero Survival Pack