Download as PDF, PPTX
Uploaded byZhenxiao Luo
Presto @ Uber Hadoop summit2017
This document discusses Presto, an interactive SQL query engine for big data. It describes how Presto is optimized to quickly query data stored in Parquet format at Uber. Key optimizations for Parquet include nested column pruning, columnar reads, predicate pushdown, dictionary pushdown, and lazy reads. Benchmark results show these optimizations improve Presto query performance. The document also provides an overview of Uber's analytics infrastructure, applications of Presto, and ongoing work to further optimize Presto and Hadoop.
More Related Content
Similar to Presto @ Uber Hadoop summit2017
Presto @ Uber Hadoop summit2017
- 1. Zhenxiao Luo Software Engineer@ Uber Even Faster: When Presto Meets Parquet @ Uber
- 2. Mission Uber Business Highlights AnalyticsInfrastructure @ Uber Presto Interactive SQL engine for Big Data Parquet Columnar Storage for Big Data Parquet Optimizations for Presto Ongoing Work Agenda
- 3. Transportation as reliableas running water, everywhere, for everyone Uber Mission
- 4. Uber Stats 6 Continents 73 Countries 450 Cities 12,000 Employees 10+ Million Avg.Trips/Day 40+ Million MAU Riders 1.5+ Million MAU Drivers
- 5. Kafka Analytics Infrastructure @Uber Schemaless MySQL, Postgres Vertica Streamio Raw Data Raw Tables Sqoop Reports Hadoop Hive Presto Spark Notebook Ad Hoc Queries Real Time Applications Machine Learning Jobs Business Intelligence Jobs Cluster Management All-Active Observability Security Vertica Samza Pinot Flink MemSQL Modeled Tables Streaming Warehouse Real-time
- 6. Parquet @ Uber RawTables â—Ź No preprocessing â—Ź Highly nested â—Ź ~30 minutes ingestion latency â—Ź Huge tables Modeled Tables â—Ź Preprocessing via Hive ETL â—Ź Flattened â—Ź ~12 hours ingestion latency
- 7. Scale of Presto@ Uber â—Ź 2 clusters â—‹ Application cluster â– Hundreds of machines â– 100K queries per day â– P90: 30s â—‹ Ad hoc cluster â– Hundreds of machines â– 20K queries per day â– P90: 60s â—Ź Access to both raw and model tables â—‹ 5 petabytes of data â—Ź Total 120K+ queries per day
- 8. â—Ź Marketplace pricing â—‹Real-time driver incentives â—Ź Communication platform â—‹ Driver quality and action platform â—‹ Rider/driver cohorting â—‹ Ops, comms, & marketing â—Ź Growth marketing â—‹ BI dashboard for growth marketing â—Ź Data science â—‹ Exploratory analytics using notebooks â—Ź Data quality â—‹ Freshness and quality check â—Ź Ad hoc queries Applications of Presto @ Uber
- 9. What is Presto:Interactive SQL Engine for Big Data Interactive query speeds Horizontally scalable ANSI SQL Battle-tested by Facebook, Uber, & Netflix Completely open source Access to petabytes of data in the Hadoop data lake
- 10.
- 11. Why Presto isFast â—Ź Data in memory during execution â—Ź Pipelining and streaming â—Ź Columnar storage & execution â—Ź Bytecode generation â—‹ Inline virtual function calls â—‹ Inline constants â—‹ Rewrite inner loops â—‹ Rewrite type-specific branches
- 12. Resource Management â—Ź Prestohas its own resource manager â—‹ Not on YARN â—‹ Not on Mesos â—Ź CPU Management â—‹ Priority queues â—‹ Short running queries higher priority â—Ź Memory Management â—‹ Max memory per query per node â—‹ If query exceeds max memory limit, query fails â—‹ No OutOfMemory in Presto process
- 13. Limitations â—Ź No faulttolerance â—Ź Joins do not fit in memory â—‹ Query fails â—‹ No OutOfMemory in Presto process â—‹ Try it on Hive â—Ź Coordinator is a single point of failure
- 14. No Need toCopy Data: Presto Connectors
- 15. Parquet: Columnar Storagefor Big Data
- 16. Parquet Optimizations forPresto Example Query: SELECT base.driver_uuid FROM hdrone.mezzanine_trips WHERE datestr = '2017-03-02' AND base.city_id in (12) Data: â—Ź Up to 15 levels of Nesting â—Ź Up to 80 fields inside each Struct â—Ź Fields are added/deleted/updated inside Struct
- 17.
- 18.
- 19.
- 20.
- 21.
- 22.
- 23.
- 24. Presto Ongoing Work â—ŹGeoSpatial query optimization â—Ź Presto Elasticsearch Connector â—Ź Multi-tenancy Support â—Ź All Active Presto Cross Data Centers â—Ź Authentication and Authorization â—Ź High Available Coordinator
- 25. Hadoop Infrastructure &Analytics â—Ź HDFS Erasure Encoding â—Ź HDFS Tiered Storage â—Ź All Active Hadoop Cross Data Centers â—Ź Hive On Spark â—Ź Spark â—Ź Data Visualization
- 26. Thank you Proprietary andconfidential © 2016 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. We are Hiring https://www.uber.com/careers/list/27366/ Send resumes to: abhik@uber.com or luoz@uber.com Interested in learning more about Uber Eng? Eng.uber.com Follow us on Twitter: @UberEng