[Strata NYC 2019] Turning big data into knowledge: Managing metadata and data relationships at Uber's scale
•
3 j'aime•445 vues
Discover how Uber thinks about building big data knowledge platforms to allow teams to discover, manage, and govern entities. Explore how to build an extensible metadata management platform and infrastructure to democratize data at Uber's scale
Recommandé
Recommandé
Contenu connexe
Tendances
Tendances (20)
Similaire à [Strata NYC 2019] Turning big data into knowledge: Managing metadata and data relationships at Uber's scale
Similaire à [Strata NYC 2019] Turning big data into knowledge: Managing metadata and data relationships at Uber's scale (20)
Dernier
Dernier (20)
[Strata NYC 2019] Turning big data into knowledge: Managing metadata and data relationships at Uber's scale
- 1. Turning Big Data into Knowledge September 25, 2019 Kaan Onuk, Luyao Li, Atul Gupte
- 2. Hi, welcome!
- 3. ● Engineer turned Product Manager ● Previously: building FarmVille & the mobile advertising platform @ Zynga ● Currently: Product Manager on the Data Platform team building data science, data knowledge, and interactive analytics platforms About me Atul Gupte Product Manager
- 4. What we’ll talk about today Data landscape at Uber Our journey since 2016 Metadata management through Databook dtbData relationships through Lineage lin
- 5. We ignite opportunity by setting the world in motion 15M Trips/Day 700+ Cities 100M Monthly Users
- 6. Data informs every decision at the company
- 7. Daily Uber trips powered by ML Millions Messages processed by Kafka 2T Queries across Hive, Vertica and Presto 1M Data ingested into HDFS 150TB How Big is our Big Data?
- 8. Data Platform Team Move the world with global data, local insights, and intelligent decisions.
- 9. Data Infrastructure Data Platform DataTools Data Lake Logging Stream Data Modelers Data Consumers ... Trips Users Data Engineers Overview of Data at Uber Data Scientists
- 10. Raw ~10,000 Curated ~100 Derived >100,000 Data LakeSources Usage WAU 8,000+ Queries 1M/day Pipelines Thousands Metrics Thousands Experiments Thousands ML models 10s of thousands Self Serve & Open Platform Use Cases Eng ETA, surge, safety DS incentives, churn, pickup Ops driver onboarding, eats cash, partner data sharing Compliance ops metrics, city Challenges compounded by the scale of Data Data produced by Mobile users 100s of millions Events Trillions/day
- 11. What are users looking to do? What data exists? How does it look? Who’s using it? What happens if I change it? How can I adapt when this data changes?
- 12. Discover Understand Trust 3+ hours week
- 13. 8% Time wasted every week $$$M Cost to company
- 15. We power data fluency to help Uber make confident, data-driven decisions Any and all users can access and use datasets with ease Users trust our data because it meets their expectations Users access appropriate data, through compliant means
- 19. Discover
- 20. Late 2016 ● Indexed small amount of data ○ Offline analytics systems ● Datasets only - no other data entities ● Catalogued basic information about datasets
- 21. Late 2016 Novice Neville Data Scientists Software Engineers ML Researchers New to Uber Requires help finding data Relies on George for basic tasks Manager Michelle General Managers Product Managers City Operations CXOs & other executives Interface w/regulators & customers Meet critical deadlines Deliver reports and insights Genius George Data Scientists Software Engineers ML Researchers Built underlying systems Tribal knowledge champion De-facto knowledge bank
- 22. Late 2016
- 23. 2017 HQ Non-HQ Rideshare Eats Freight ATG Elevate Support NLP models for support tickets Safety Trip classification Uber Eats Restaurant recommendations Operations LTV models
- 24. 2017 Data Scientists Software Engineers ML/AI Researchers Advanced SQL Advanced Statistics Scala/Spark, Python/R Data Modeling Inventor Ivan Marketing Managers Entry-level Analysts General Managers Product Managers Limited SQL Spreadsheets Reliant Rebecca City Operations Regional Managers Advanced SQL Spreadsheets Dashboarding Monitoring Matt Operations Managers Data Analysts Product Analysts Advanced SQL Spreadsheets Limited Statistics Limited Python/R Analyst Anna
- 25. 2017
- 26. 2018Cumulativefunctionality Time Low internal quality High internal quality Delivers more rapidly + cheaply later
- 27. ● Users care about a variety of data assets ○ Datasets, dashboards, metrics, etc. ● Users want a holistic view of everything that exists about their data ○ Ownership ○ Schemas 2018
- 28. 2018
- 29. 2018 ● Data quality and health are key concerns ● Table usage information is valuable ● Operational and regulatory environment is growing more complex ○ GDPR ○ Access control & audits
- 30. 2019
- 31. 2019 ● Unified interface highlighting relevant metadata ○ Ownership & usage ○ Schema and stats ○ Quality & health signals ○ Lineage
- 32. 2019 ● Advanced metadata management ○ Automated ingestion ○ Automated classification ○ Simplified controls for data owners 60+ Types of metadata
- 34. ● Manages Data Lineage team under Data Platform ● Earlier: Senior Software Engineer II at Uber About me Luyao Li Tech Lead Manager
- 35. Data Lineage
- 36. What is data lineage? ● Where is the data from ● Where it’s been ● How it’s being transformed
- 37. “I’m no longer responsible for this table, please ask team X” “This is an upstream problem, we can’t fix it” “Please ask the table owner” “How do I find the pipeline owner?” Multiple days Why does it matter?
- 38. Applications Data Freshness Data Chargebacks Anomaly Detection Compliance
- 40. 10,000-foot view
- 41. 1,000-foot view
- 42. Lessons learned High quality data is essential for success Always be customer obsessed Magical search has a huge impact on usability Make big, bold bets
- 43. What’s next ● Column-level lineage ● Self-diagnostic and reporting ● Self-serve onboarding ● Recommendation
- 44. / Manages the Metadata Platform team within Big Data / Previously - Senior Software Engineer / Tech Lead for Data Discovery & Data Privacy @ Uber About me Kaan Onuk Engineering Manager
- 45. / What is metadata? / Why does metadata matter?
- 47. Uber’s massive data holds deep hidden insights. Metadata helps to surface them.
- 48. Metadata drives data productivity by making data easy to discover, understand, and govern.
- 49. / Metadata Sources / Metadata Registry / Metadata Collection / Metadata Storage / Data Model
- 52. Metadata Collection Pull model Push model ○ Crawler (periodic) e.g. sample data, stats ○ Event-based (Event Listeners) e.g. data quality ○ Automated e.g. data retention policies ○ Crowdsource e.g. table descriptions
- 53. Storage ● Hive for analytical queries and audit purposes ● Kafka to capture metadata changes ● MySQL for persistent storage ● Redis for cache to support low latency & high throughput ● Search functionality powers various internal platform including Databook for data discovery
- 54. Metadata Store: Data Model Requirements 1. Discovery 2. Cluster-specific & agnostic metadata 4. Flexibility on onboarding new entities 3. Easy metadata type creation
- 56. Metadata Management (2019) 1. Easy onboarding 2. Derived metadata through relationships 3. Efficient metadata retrieval
- 57. Key Takeaways 1. Centralized Metastore: Datasets + Artifacts 2. Metadata Registry: Taxonomy / Metadata scheme 3. Metadata Collection: Choose the right approach 4. Data Model: Leverage metadata relationships
- 58. Next Steps Metadata Management Innovate - Personalization to improve discovery - Graph traversal optimizations Automate - Human-in-the-loop AI Establish trust & accountability - More integrations with data infra - Very high qps & low latency Accessible but secure foundation - Fully self-served, ontology-based metadata management
- 60. Proprietary and confidential © 2018 Uber Technologies, Inc. All rights reserved. No part of this document may be reproduced or utilized in any form or by any means, electronic or mechanical, including photocopying, recording, or by any information storage or retrieval systems, without permission in writing from Uber. This document is intended only for the use of the individual or entity to whom it is addressed and contains information that is privileged, confidential or otherwise exempt from disclosure under applicable law. All recipients of this document are notified that the information contained herein includes proprietary and confidential information of Uber, and recipient may not make use of, disseminate, or in any way disclose this document or any of the enclosed information to any person other than employees of addressee to the extent necessary for consultations with authorized personnel of Uber. Thank you! kaan@uber.com