20-minute talk for the Knowledge Graph Conference, May 2019

1. Building an Enterprise
Knowledge
Graph @Uber:
Lessons from Reality
Joshua Shinavier, PhD
Knowledge Graph Conference
May 8th
, 2019
...

2. The future
Half empty
Half full
The present
Knowledge @Uber

3. ● Uber is an ideal proving ground for an enterprise knowledge graph (EKG)
● 200k managed data sets
● Billions and billions of trips served
○ Low thousands of new entities per second
○ Totally doable!
● Even more sensor data
○ Use cases for graph stream processing
● Genuine need for knowledge and real-time inference
Knowledge @Uber

4. EKG hierarchy of needs

6. ● Real data is messy

7. ● Real data is messy
● We are not all ontologists

8. ● Real data is messy
● We are not all ontologists
● Good enough does not scale

9. ● Real data is messy
● We are not all ontologists
● Good enough does not scale
● Beware of the hype cycle

10. ● Real data is messy
● We are not all ontologists
● Good enough does not scale
● Beware of the hype cycle
● RDF is a hard sell

11. ● Real data is messy
● We are not all ontologists
● Good enough does not scale
● Beware of the hype cycle
● RDF is a hard sell
● Property Graphs are not enough

12. ● Use and promote standards

13. ● Use and promote standards
● Invest in shared vocabulary

14. ● Use and promote standards
● Invest in shared vocabulary
● Fit the tooling to the infrastructure

15. ● Use and promote standards
● Invest in shared vocabulary
● Fit the tooling to the infrastructure
● Fit the data model to the data

16. ● Use and promote standards
● Invest in shared vocabulary
● Fit the tooling to the infrastructure
● Fit the data model to the data
● Budget for “other stuff”

17. ● Use and promote standards
● Invest in shared vocabulary
● Fit the tooling to the infrastructure
● Fit the data model to the data
● Budget for “other stuff”
● Collaborate early and often

18. Risk & Safety Knowledge Graph
This slide intentionally left blank to save entropy.
UBER KNOWLEDGE GRAPH

19. ● Controlled vocabularies for all of Uber
○ Basic type aliases
○ Structured types for geospatial data, sensor data, money, etc. etc.
○ Entities and relationships (User, Vehicle, Trip, etc.)
○ Metadata vocabularies
● Elevates domain-specific RPC and storage schemas to ontologies
● Tooling carries schemas between data representation languages
○ Protobuf, Thrift, Avro, RDF, PG, etc.
Data Standardization

20. ● Hundreds of thousands of structured datasets at Uber
● Data protections and user trust
○ GDPR and other regulations, Uber’s own data policies
○ What kind of user data? Where is it?
○ Heroic numbers of manual annotations
■ Limited expressivity, limited guarantees
■ Inference is required
● Two birds: in annotating datasets, standardize and compose schemas
○ Now we have a true global knowledge graph
○ Investigating efficient reasoning and “No ETL” solutions
Metadata graph

21. ● Common data model for RPC, storage, and KR at Uber
● In progress: alignment with the Property Graph Schema Working Group
● In progress: “Universal structure” of TinkerPop4
Algebraic Property Graphs

22. ● Real data is messy
● We are not all ontologists
● Good enough does not scale
● Beware of the hype cycle
● RDF is a hard sell
● The Property Graph is not enough
● Use and promote standards
● Invest in shared vocabulary
● Fit the tooling to the infrastructure
● Fit the data model to the data
● Budget for “other stuff”
● Collaborate early and often

23. joshsh@uber.com
Thanks