Running Databases Efficiently at Scale - Case Study of Flipka...
summary
Flipkart moves stateful databases onto Kubernetes using Local PVs with OpenEBS + LVM to get 2x–4x better IOPS/throughput and half the cost (on 4PB SSD + 15PB HDD across 1000 nodes and 3 clusters). Contributions add LVM dynamic provisioning via CSI and a CSI storage capacity integration so the scheduler sees per-node disk. Node maintenance is handled by a CR-based maintenance workflow and cgroups enforce IOPS limits; outcomes: 3x ops speedup and 2x–8x cost cuts, but no online volume expansion and CSI capacity is eventually consistent.
AI-written summary · about
chapters (15)
- 0:00Flipkart's infrastructure journey
- 1:03Scale and storage footprint
- 3:33Kubernetes for stateful workloads
- 6:04Local vs network storage
- 8:57Cost and performance findings
- 9:18Local storage challenges
- 11:28Dynamic local volumes
- 15:40Pod scheduling with local storage
- 18:34Noisy neighbor problem
- 21:44Day-two operations
- 25:19Distributed databases success
- 27:51Local storage trade-offs
- 29:34Key takeaways
- 31:19Q&A part one
- 41:03Q&A part two
related talks
29mFlipkart: Scaling Aerospike at Flipkart with Kubernetes
27mPods on Steroids - A Case Study of Migrating Flipkart.com to K8s on Bare Metal
21mCase Study: FOSS (Apache Hbase) at Flipkart
38mUsing OpenEBS successfully for remote working and to reduce costs
35meBay Geo-Distributed Database on Kubernetes [A]
49mHow Flipkart runs HBase across zones and clouds seamlessly
30mStateful Superpowers: Explore High Performa...
01h 02mScaling One of the Largest On-prem Managed MySQL Fleets in India