Uber achieves millions $ of savings managing disk IO across HDFS cluster
summary
Company achieves ~45% storage unit cost reduction by deploying 16TB HDDs across HDFS and rebalancing IO so average cluster IOPS per TB stays even, yielding millions in savings. Engineers add same-disk tiering (per-directory capacity ratios so HOT/COLD blocks share a single drive), enable hard-link moves to avoid copy traffic, and implement client-side fast-switch read plus namenode deprioritization of slow datanodes. Monitoring, mover tuning and production tests show unified IO and controlled tail latency.
AI-written summary · about
chapters (12)
- 0:00Introduction to Uber's data infrastructure
- 3:18HDFS scale and storage challenges
- 6:51Goals for cost-efficient storage tiering
- 9:20Current multi-tier HDFS architecture
- 13:37High-density storage hardware trends
- 17:50Disk partitioning for IO balancing
- 21:56HDFS storage directory configuration
- 24:25Performance impact of 16TB HDDs
- 26:33Data temperature analysis for tiering
- 27:57Monitoring and metrics for HDFS
- 33:36Client-side read improvements
- 37:12Q&A session
related talks
25mGreenOps at Scale: The 50TB Memory Diet for a Petabyte Platform
17mBoosting Cost Efficiency by Cost Aware Architecture
38mScaling IO-bound application in the cloud
57mDriving innovation and insight with storage
33mContainerizing & Orchestrating Hadoop Infrastructure at Scale
42mBillion-Scale Triumph: Innovating at the Edge of Cloud Data
24mDynamic Big Data in the Cloud with Kubernetes
20mLogging in the Age of Cost-Cutting