Data @Scale 2024
summary
Meta reports a 3x speedup moving Presto workloads to a C++ execution library called Vu/Velox, introduces a Nimble file format and an MTA V1 accelerator while planning 600,000 GPUs; Snowflake describes training Arctic on ~3.5 trillion tokens using Snowflake for preprocessing; Razorpay achieves 99.99% availability, thwarts 542 DDoS attacks and cuts costs 60% as TPV grows 70%; incident teams use LLM assistants to find root causes ~42% of the time within minutes, and BML/Beam claims 3x GPU efficiency and up to 60% fleet utilization improvements.
AI-written summary · about
chapters (12)
- 0:00Welcome to Data @Scale 2024
- 23:26AI First Data Infrastructure at Meta
- 32:57Generative AI challenges
- 42:29Responsible AI impact
- 1:10:40Evolving Data Management at Meta
- 1:54:06Razorpay’s payment infrastructure
- 2:37:46Generative AI for incidents
- 3:10:40Meta’s Data Graph for privacy
- 3:53:58Big Data infrastructure revisited
- 4:23:59Generative AI in AWS analytics
- 4:53:58Apache Beam ML for AI
- 5:10:00Fireside Chat on AI infrastructure
related talks
14mBig Data Processing with GPUs
22mPresto for Your Fast, Open SQL Lakehouse
28mInnovations in AWS analytics: Data processing
39mUnlock the potential of your data: Optimize your data store architecture...
42mMulti-GPU Accelerated Processing of Time-Series Data of Huge Academic
43mArchitecting for Speed & Scale
30mIntroducing Spark Cyclone Accelerating Spark with...
27mEnterprise AI at Scale: Success Stories and Insights