Zeus is an efficient, highly scalable and distributed shuffle as a service which is powering all Data processing (Spark and Hive) at Uber. Uber runs one of the largest Spark and Hive clusters on top of YARN in industry which leads to many issues such as hardware failures (Burn out Disks), reliability and scalability challenges.

1. Edit or delete footer text in Master ipsandella doloreium dem isciame ndaestia nessed
quibus aut hiligenet ut ea debisci eturiate poresti vid min core, vercidigent.
Zeus: Uber’s Highly Scalable and
Distributed Shuffle as a Service
Mayank Bansal, Data Infra, Uber
Bo Yang, Data Infra, Uber
Igniting opportunity by setting the world in motion

2. 15 billion trips
18M trips per day
6 continents, 69 countries and 10,000 cities
103M active monthly users
5M active drivers
22,000 employees worldwide
3,700 developers worldwide
2

3. Data and ML Use Cases at Uber
○ Uber Eats
○ ETAs
○ Self-Driving Vehicles
○ Customer Support
○ Driver/Rider Match
○ Personalization
○ Demand Modeling
○ Dynamic Pricing
○ Forecasting
○ Maps
○ Fraud
○ Anomaly Detection
○ Capacity Planning
○ And many more...

4. Data and ML at Uber - ETAs
○ ETAs are core to the Uber customer experience
○ ETAs used by myriad internal systems
○ ETA are generated by route-based algorithms
○ ML models predict the route-based ETA error
○ Uber uses the predicted error to correct the
ETA
○ ETAs now dramatically more accurate

5. Data and ML at Uber - Driver/Rider Match
○ Optimize matchings of riders and drivers
on the Uber platform
○ Predict if open rider app will make trip
request

6. Data and ML at Uber - Eats
○ Models used for
○ Ranking of restaurants and
dishes
○ Delivery times
○ Search ranking
○ 100s of ML models called to
render Eats homepage

7. Data and ML at Uber - Self-Driving Vehicles

8. Edit or delete footer text in Master ipsandella doloreium dem isciame ndaestia nessed
quibus aut hiligenet ut ea debisci eturiate poresti vid min core, vercidigent.
Uber’s Data Stack
Mobile App Events
Device Telemetry
Micro-Service Events
Database Events
3rd Party Feeds
Bulk Uploads
Incremental
Ingestion
Kafka
Realtime, Pre-Aggregated
(AthenaX)
Ad hoc, Interactive
(Presto, Vertica)
Complex, Batch
(Hive)
Dashboards
(Summary, Dashbuilder)
Ad hoc Query
(QueryBuilder)
Data Preparation
(Piper, uWorc)
BI Tools
(Tableau, DSW)
Stream Processing
(Flink)
Batch Processing
(Spark, Tez, Map Reduce)
Compute Fabric (YARN / Mesos + Peloton)
Data Analytics Tools
In-memory
(Pinot,
AresDB)
Hot
(HDFS)
Warm
(HDFS)
Archival
(Cloud)
Query Engines
Data Processing Engines
Tiered Data Lake

9. Edit or delete footer text in Master ipsandella doloreium dem isciame ndaestia nessed
quibus aut hiligenet ut ea debisci eturiate poresti vid min core, vercidigent.
Uber’s ML Stack - Michelangelo
Kafka
Compute Fabric (YARN / Peloton+Mesos)
Data Analytics Tools
Query EnginesStream
Processing
(Flink)
Batch
Processing
(Hive,
Spark, Tez)
Data
Preparation
Jupyter
Notebook
Spark
Magic
Prototype
Tensorflow
Training
Pytorch
XGBoost
SparkML
Feature
Store
Model
Store
Metrics
Store
DataLake
(HDFS)
Inference
Realtime
Prediction
Service
Batch
Prediction
Jobs

10. Apache Spark
@Uber
Image Source: www.mindproject.io

11. Edit or delete footer text in Master ipsandella doloreium dem isciame ndaestia nessed
quibus aut hiligenet ut ea debisci eturiate poresti vid min core, vercidigent.
○ Apache Spark is the primary analytics execution
engine teams at Uber use
○ At Uber, 95% batch and ML jobs run on Spark
○ We run Spark on YARN and Peloton/Mesos
○ We use external shuffle service for the
shuffle data
Apache Spark @ Uber
11
* Apache Hadoop,, Spark, and Mesos logos are either registered trademarks or trademarks of the Apache
Software Foundation in the United States and/or other countries. No endorsement by The Apache
Software Foundation is implied by the use of these marks. TensorFlow and the TensorFlow logo are
trademarks of Google Inc. Redis is a trademark of Redis Labs Ltd. Any rights therein are reserved to Redis
Labs Ltd. Any use by Uber Technologies is for referential purposes only and does not indicate any
sponsorship, endorsement or affiliation between Redis and Uber Technologies.

12. 12
How Does Apache Spark Shuffle Service Work?

13. Limitations of Apache Spark Shuffle Service
13
● SSD wearing out Issues
● Reliability
● Kubernetes dynamic allocation
● Collocation

14. Different Approaches
14
● Shuffle manager to external storage
○ Synchronous writes
■ NFS
● 2X slow
■ HDFS
● 5X slow

15. Different Approaches
15
● Shuffle manager to external storage
○ Semi-asynchronous writes
■ HDFS
● 4x slow

16. Different Approaches
16
● Remote Shuffle Service
○ Streaming writes to HDFS
■ 1.5x slower than writing to local storage
○ Streaming Writes to Local
■ ~Same Performance like external shuffle service

17. Remote Shuffle Service
17
● Remote Shuffle Service
○ Streaming Writes to Local Storage
■ Changed Mapreduce paradigm
■ Record Stream -> Shuffle Server -> Disk
■ No temporary spill files in executor side

18. Architecture - Remote Spark Shuffle Service
18

19. Deep Dive
Image Source: www.mindproject.io

20. Design Principles
20
● Scale out horizontally
○ Each server instance works independently
○ Avoid centralized state/storage
● Tackle network latency
○ Reduce waiting times for server response
○ Stream data
● Performance optimization
○ Most Spark Apps optimized for similar performance
○ Rely on YARN/Apache Spark retry for failure recovery

21. Scale Out
Tackle Network Latency
Performance Optimization
21

22. Horizontal Scalable
22
● Spark applications share/use different shuffle servers
● No shared state among shuffle servers
● More shuffle servers to scale out

23. Shuffle Server Distribution
23
● Mappers: m=4
● Reducers: r=5
● Shuffle Servers: s=3

24. Shuffle Server Distribution in General
24
● Mappers: m
● Reducers: r
● Shuffle Servers: s
● Network Connections
○ Mappers: m*s connections
○ Reducers: r connections

25. Scale Out
Tackle Network Latency
Performance Optimization
25

26. Server Implementation
26
● Use Netty
○ High performance asynchronous server framework
● Two thread groups
○ Group 1: Accept new socket connection
○ Group 2: Read socket data
○ Thread groups not block each other
● Binary network protocol
○ Efficient encoding/compression

27. Direct Write/Read on Disk File
27
● Write to OS file directly
○ No application level buffering
● Zero copy
○ Transfer data from disk file to shuffle reader without user space
memory
● Sequential write/read
○ No random disk IO

28. Client Side Compression
28
● Shuffle client compress/decompress data
● Reduce network transport data size
● Reduce CPU usage on shuffle server
● Support client side encryption
○ Encryption key inside each application
○ Encryption key not distributed to shuffle server

29. Parallel Serialization and Network IO
29
● Shuffle data serialization takes time
● Serialization in executor thread
● Network IO in another thread

30. Connection Pool
30
● Socket connect latency is not trivial
● Reuse client/server connections

31. Scale Out
Tackle Network Latency
Performance Optimization
31

32. Asynchronous Shuffle Data Commit
32
● Map task
○ Stream data to server
○ Not wait for response
● Server flushes (commits) data asynchronously
● Reduce task queries data availability when fetching data

33. Fault Tolerance
33

34. Shuffle Server Discovery/Health Check
34
● ZooKeeper as Server
Registry

35. Data Replica
35
● Server Replication Group
● Duplicate Write in Parallel
● Read from Single Server,
switch to another server on
failure

36. Local State Flush
36
● Local state persistence in batch
○ Avoid flushing state for each map task
○ Flush when shuffle stage finishes
● Client not waiting for server side state flush

37. Production Status
37

38. Compatible with Open Source Apache Spark
38
● Shuffle Manager Plugin
○ spark.shuffle.manager=
org.apache.spark.shuffle.RssShuffleManager
● MapStatus / MapOutputTracker
○ Embed remote shuffle service related data inside MapStatus
○ Query MapOutputTracker to retrieve needed information

39. Metrics/Monitoring
39
● Uber’s open source M3 metrics library
● Important metrics
○ Network connections
○ File descriptors
○ Disk utilization

40. Test Strategy
40
● Unit Test
● Stress/Random Test
● Production Query Sampling

41. Remote Spark Shuffle Service - Production Status
41
● In production in last 8+ months for
YARN
● Thousand’s of application running
every day
● Job latencies are on par with
external shuffle
● Open sourcing it soon!

42. Roadmap
42
● Support all Spark workloads including HiveOnSpark
● Multi-tenancy (quota)
● Load balancing
● Integrate with incoming Spark shuffle metadata APIs

43. Proprietary and confidential © 2020 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.
4
3
Thank you !!!