Presto Apache BigData 2017
2 likes482 views
Zhenxiao Luo presented on optimizations made to Presto to improve its performance when querying Parquet files at Uber. Presto is an interactive SQL query engine used at Uber to query raw and modeled data stored in Hadoop. Parquet is a columnar storage format used at Uber. The optimizations made to Presto's Parquet reader include nested column pruning, columnar reads, predicate pushdown, dictionary pushdown, and lazy reads. These optimizations resulted in Presto query performance improvements of up to 20x for Parquet workloads at Uber's scale of petabytes of data and tens of thousands of queries per day.
Presto Apache BigData 2017
- 1. Zhenxiao Luo Software Engineer @ Uber Even Faster: When Presto Meets Parquet @ Uber
- 2. Mission Uber Business Highlights Analytics Infrastructure @ Uber Presto Interactive SQL engine for Big Data Parquet Columnar Storage for Big Data Parquet Optimizations for Presto Ongoing Work Agenda
- 3. Transportation as reliable as 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 Raw Tables ● No preprocessing ● Highly nested ● ~30 minutes ingestion latency ● Huge tables Modeled Tables ● Preprocessing via Hive ETL ● Flattened ● ~12 hours ingestion latency
- 7. Scale of Presto @ Uber ● 2 clusters ○ Application cluster ■ Hundreds of machines ■ 100K queries per day ■ P90: 30s ○ Ad hoc cluster ■ Hundreds of machines ■ 20K queries per day ■ P90: 60s ● Access to both raw and model tables ○ 5 petabytes of data ● Total 120K+ queries per day
- 8. ● Marketplace pricing ○ Real-time driver incentives ● Communication platform ○ Driver quality and action platform ○ Rider/driver cohorting ○ Ops, comms, & marketing ● Growth marketing ○ BI dashboard for growth marketing ● Data science ○ Exploratory analytics using notebooks ● Data quality ○ Freshness and quality check ● 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. How Presto Works
- 11. Why Presto is Fast ● Data in memory during execution ● Pipelining and streaming ● Columnar storage & execution ● Bytecode generation ○ Inline virtual function calls ○ Inline constants ○ Rewrite inner loops ○ Rewrite type-specific branches
- 12. Resource Management ● Presto has its own resource manager ○ Not on YARN ○ Not on Mesos ● CPU Management ○ Priority queues ○ Short running queries higher priority ● Memory Management ○ Max memory per query per node ○ If query exceeds max memory limit, query fails ○ No OutOfMemory in Presto process
- 13. Limitations ● No fault tolerance ● Joins do not fit in memory ○ Query fails ○ No OutOfMemory in Presto process ○ Try it on Hive ● Coordinator is a single point of failure
- 15. Parquet: Columnar Storage for Big Data
- 16. Parquet Optimizations for Presto Example Query: SELECT base.driver_uuid FROM hdrone.mezzanine_trips WHERE datestr = '2017-03-02' AND base.city_id in (12) Data: ● Up to 15 levels of Nesting ● Up to 80 fields inside each Struct ● Fields are added/deleted/updated inside Struct
- 19. Columnar Reads
- 22. Lazy Reads
- 24. Ongoing Work ● Multi-tenancy support ● High availability for coordinator ● Geospatial optimization ● Authentication & authorization
- 25. Thank you Proprietary and confidential © 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