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.

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

14. Presto Connectors

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

17. Old Parquet Reader

18. Nested Column Pruning

19. Columnar Reads

20. Predicate Pushdown

21. Dictionary Pushdown

22. Lazy Reads

23. Benchmarking Results

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