注册 / 登录

Real Time Analytics on Deep Learning: when Tensorflow meets Presto at Uber

分会场:  人工智能/AI驱动/AI实践

 

案例来源 :

案例讲师

罗震霄

Uber Engineering Manager, Big Data Department

Zhenxiao is an Engineering Manager at Uber, where he runs Interactive SQL engine, columnar format, and caching projects for big data. Previously, Zhenxiao led the development and operations of SQL engine at Netflix and worked on big data related projects at Facebook, Cloudera, and Vertica. Zhenxiao holds a master’s degree from the University of Wisconsin-Madison and a bachelor’s degree from Fudan University.

扫描二维码分享案例

 

案例简述

 

From determining the most convenient rider pickup points to predicting the fastest routes, Uber uses machine learning and data-driven analytics to create seamless trip experiences.

Inside Uber, big data and machine learning are spread everywhere. Analysts and Engineers would like to run real time analytics with deep learning models. While, copy data from one source to another is pretty expensive. It is challenging to support real time analytics with deep learning.

This talk will share Uber’s engineering effort, supporting real time analytics with deep learning on the fly, without any data copy. We will start with our big data and deep learning infrastructure, specifically Tensorflow, Hadoop, and Presto. Then we will talk about how Uber used Tensorflow as deep learning engine, and Presto as the interactive SQL engine. We will focus on how Uber built Presto Tensorflow Connector from scratch, to support real time analytics on deep learning. Finally, we will share our production experien

 

案例目标

 

Big data and deep learning are used by companies to make business decisions, give user recommendations, and analyze experiments across all data sources. While, data systems and deep learning systems are managed by different engineering/operation departments, and it is costly to copy data from one system to another. This makes it challenging to run analytics at company’s global point of view, and hurts deep learning models’ accuracy, due to lack of necessary training data. At Uber, big data system and deep learning system are two independent systems. We were facing many engineering and management challenges, when customers would like to run real time analytics on deep learning data, or scientists would like to train models on big data. The time cost is huge when copying data between these two systems. We’d like to share Uber’s engineering experience integrating big data and deep learning systems together, by building the Presto TensorFlow Connector from scratch. In our new architecture, customers could use Presto to query TensorFlow data on the fly, without the need to copy data, at the same time, data scientists could train models directly on Hadoop data. The Presto TensorFlow Connector project greatly improved data scientists’ productivity, and became one of the most popular projects. It saved all the resource of data copy, e.g. hundreds of terabytes of disk space, and millions of CPU cycles.

 

成功(或教训)要点

 

1、Choose the right technology Use open source technology? Or buy proprietary software? Reliability, Scalability, Performance We need engineers to own the technology 2、Choose the right people Why engineers are working here? Why engineers would like to collaborate? We need to make environments for engineers to collaborate 3、Make the right decision Who should make decision? What’s the procedure of making decision? Let builders build 4、On the right trend What does our architecture look like in 1 year, 2 years, 5 years? Do we need to build these things? Shall we go cloud? Why do we need on prem data centers?

 

案例ROI分析

 

Horovod improves TensorFlow Deep Learning throughput by 2X Horovod supports tens of Deep Learning models in production Presto TensorFlow Connector improves SQL performance by 10X Presto TensorFlow Connector supports hundreds of production jobs daily

 

案例启示

 

-Where shall we play with our data? Using public cloud, or building on premises data center -public cloud provides a whole spectrum of services, with simplicity. At early stage, startups could leverage public cloud to build whole data pipeline -public cloud provides elasticity. Using public cloud, we could immediately get 2X more hardware, with all software stack support -although building on prem data center is time consuming, it is much cheaper compared with public cloud -building your own data center also means, you own the software. The company will no longer be locked by any vendor, and could provide better support with more flexibility -when your startup is at its early

 

案例在团队中的意义

 

Share big data and deep learning architecture experience with software engineers, data scientists, and machine learning engineers

 

领取大会PPT

我要参会

大会全套演讲PPT

立即领取

大会即将开幕,点击抢票!

我要参会