Manage Scaling, While Keeping Costs Under Control. Challenge Accepted.
summary
Relativity runs 17 clusters in 10 regions on Azure, manages 6,000 VMs and 8 PB of data, and moves everything into Kubernetes. Scaling pain points include heavy variance between tiny microservices and multi‑core, 100+ GB batch jobs, Azure storage I/O throttling that produces "zombie" nodes and mass pod evictions from emptyDir abuse. Remedies: internal PaaS, multiple node‑pool sizes, reduce pods-per-node, use descheduler, node problem detector + drano, migrate teams to managed disks and PV/PVCs with Open Policy Agent enforcement, and track per-namespace/customer spend with Harness Continuous Efficiency.
AI-written summary · about
chapters (14)
- 0:00Manage scaling while keeping costs under control
- 0:55Agenda
- 1:12How does Relativity Operate Kubernetes?
- 5:06Challenges with Bin
- 6:17Auto Scaling and Bin
- 8:37Storage Throttling
- 9:34So what blew up?
- 11:14How do you put out the fire?
- 13:04Unallocated Resources
- 15:01What is the compute cost of a customer?
- 16:02Leverage Label Taxonomy
- 16:35Build or Buy
- 18:12Harness to the rescue
- 18:51Key Takeaway
related talks
15mTurning Growing Pains To Growing Wins With Azure Cosmos DB
28mScaling in Kubernetes Safely on On-Prem KaaS Across 1,300+ Clusters and 40,000+ Nodes
33mNavigating the Seas of Transformation: Migrating from EC2 to Kubernetes
57mScaling Azure Container Apps: Maximize Performance and Minimize Costs
31mAutoscaling Apps without Supersizing Costs
29mMicrosoft Azure: Operating at Hyper-Scale
36mOptimizing Resource Usage in Kubernetes
44mHow to run a global, cloud scale event for 10.000 people