ODSC WEST 2019
Cloud storage brings great flexibility in management and cost-efficiency to data scientists, but also introduces new challenges related to data accessibility and data locality for machine learning applications. For instance, when the input data is stored in a remote cloud storage like AWS S3 or Azure blob storage, direct data access is often slow and expensive; but manually moving data to the training clusters can be time-consuming, complicated and often require data engineering or ETL pipelines.
This session is designed for data scientists or data engineers who work with remote and possibly multiple data sources in hybrid or multi-cloud environments. We will guide the audience to use Alluxio to greatly simplify the data preparation in these environments, covering the following topics:
- -How to setup and create POSIX endpoint for Alluxio service to unify the file system data access to S3, HDFS and Azure blob storage
- How to run Apache Spark to read input from and write output to remote storage with Alluxio as the distributed data caching layer
- How to run TensorFlow to train models backed by accessing remote input data like access local file system.
Complete the form below to access the full overview:
.png)
Presentations
Use Alluxio to Unify Storage Systems in Suning
Suning is one of the leading commercial enterprises in China with two public companies in China and Japan respectively. It uses Alluxio to unify storage systems and manage multiple HDFS clusters.
STRATA DATA CONFERENCE LONDON 2018
JD.com is China’s largest online retailer and its biggest overall retailer, as well as the country’s biggest internet company by revenue. Currently, JD.com’s BDP platform runs more than 400,000 jobs (15+ PB) daily, on a system with more than 15,000 cluster nodes and a total capacity of 210 PB.
Alluxio, formerly Tachyon, is the world’s first system that unifies disparate storage systems at memory speed. In the big data ecosystem, Alluxio lies between computation frameworks or jobs and various kinds of storage systems. Additionally, Alluxio’s memory-centric architecture enables data access orders of magnitude faster than existing solutions.
Alluxio has run in JD.com’s production environment on 100 nodes for six months. Mao Baolong, Yiran Wu, and Yupeng Fu explain how JD.com uses Alluxio to provide support for ad hoc and real-time stream computing, using Alluxio-compatible HDFSURLs and Alluxio as a pluggable optimization component. To give just one example, one framework, JDPresto, has seen a 10x performance improvement on average. This work has also extended Alluxio and enhanced the syncing between Alluxio and HDFS for consistency.
Alluxio in MOMO: Accelerating Ad Hoc Analysis
From our friends at MOMO
MOMO, a leading pan-entertainment social platform in China, has deployed Alluxio to accelerate ad-hoc query analytics. In the course of evaluating the best fit for Alluxio in their infrastructure they conducted several performance tests to understand how ad-hoc query analytics behaved in several scenarios. These tests give real-world insight to the performance benefits Alluxio provides. The MOMO findings include:
- With Alluxio, performance was improved 3-5x over the current mode
- Even when initially reading ‘cold’ data Alluxio delivered superior performance in most cases
- Alluxio can effectively scale-out to improve performance as requirements grow