Shuffling, Fast and Slow: Scalable Analytics on Serverless Infrastructure

on YouTube ↗26m

" summary

Locus implements a hybrid shuffle that buffers hot partitions in fast cache (Redis) and spills to cheap object storage (S3), guided by a performance/cost model that picks rounds, worker size and cache allocation. It turns an S3-only shuffle that would create 10 billion to 200k requests and require 100TB to 5TB cache into a multi-round procedure, yielding 4×–500× speedups over slow-storage-only baselines, up to 59% lower resource usage, and 15-minute sorts at about $163.

AI-written summary · about

" chapters (6)

1. 0:00Introduction to locust
2. 2:14The shuffle bottleneck
3. 6:11Limitations of cloud storage
4. 11:32Locust design and strategy
5. 14:56Modeling performance and cost
6. 18:35Evaluation and discussion

" related talks

• 49mUsing serverless functions for real-time observabilityLiz Fong Jones · CPH DevFest 2024
• 32mUsing Serverless Functions for Real-time ObservabilityLiz Fong Jones, Jessica Kerr · SREcon22 Americas
• 41mUsing Serverless & ARM64 for Real-Time ObservabilityLiz Fong Jones · GOTO Chicago 2024
• 31mInfiniCache: Exploiting Ephemeral Serverless Functions to Build a Cost-Effective MemoryYue Cheng, Ao Wang, Vasily Tarasov, Lukas Rupprecht, Ali Anwar, Feng Yan, Dimitrios Skourtis, Jingyuan Zhang, Xiaolong Ma · FAST '20
• 33mNext Gen of Datastores: Hot Queries, Cold StoragePhilipp Krenn · DevConf.CZ 2023
• 16mMinFlow: High-performance and Cost-efficient Data Passing for I/O-intensive Stateful...Tao Li, Yongkun Li, Yinlong Xu, John C. S. Lui, Wenzhe Zhu · 22nd USENIX Conference on File and Storage Technologies (FAST '24)
• 32mServerless Data Processing and Machine LearningSunil Mallya · LISA18
• 41mUnlocking Faster & Efficient Data Processing w/ ServerlessUma Ramadoss, Adam Wagner · GOTO EDA Day Nashville 2023