As a Data Scientist/Engineer in Python, we focus in our work to solve problems with large amounts of data but still stay in Python. This is where we are the most effective and feel comfortable. Libraries like Pandas and NumPy provide us with efficient interfaces to deal with this data while still getting optimal performance. The main problem appears when we have to deal with systems outside of our comfort ecosystem. We need to write cumbersome and mostly slow conversion code that ingests data from there into our pipeline until we can work efficiently. Using Apache Arrow and Parquet as base technologies, we get a set of tools that eases this interaction and also brings us a huge performance improvement. As part of the talk we will show a basic problem where we take data coming from a Java application through Python into using these tools.

1. (Efficient) Data Exchange with
"Foreign" Ecosystems
Uwe Korn – QuantCo – 2nd July 2019

2. About me
• Engineering at QuantCo
• Apache {Arrow, Parquet} PMC
• Focus on Python but interact with
R, Java, SAS, …
@xhochy
@xhochy
mail@uwekorn.com
https://uwekorn.com

3. Python vs R
👊

4. Python vs R
👊

5. Python & R

6. Python & R
… & Java & Rust &
Javascript & C# & Matlab
& …

7. Do we have a problem?

8. Do we have a problem?
• Yes, there are different ecosystems!

9. Do we have a problem?
• Yes, there are different ecosystems!
• PyData
• Python / R
• Pandas / NumPy / PySpark/sparklyr / Docker

10. Do we have a problem?
• Yes, there are different ecosystems!
• PyData
• Python / R
• Pandas / NumPy / PySpark/sparklyr / Docker
• Two weeks ago: Berlin Buzzwords
• Java / Scala
• Flink / ElasticSearch / Kafka
• Scala-Spark / Kubernetes

11. Do we have a problem?
• Yes, there are different ecosystems!
• PyData
• Python / R
• Pandas / NumPy / PySpark/sparklyr / Docker
• Two weeks ago: Berlin Buzzwords
• Java / Scala
• Flink / ElasticSearch / Kafka
• Scala-Spark / Kubernetes
• SQL-based databases
• ODBC / JDBC
• Custom protocols (e.g. Postgres)

12. Why solve this?
• We build pipelines to move data
• Goal: end-to-end data products
Somewhere along the path we need to talk
• Avoid duplicate work / work on converters
• We don’t want Python vs R but use each of them where they’re best.

14. Apache Arrow at its core
• Main idea: common columnar representation of data in memory
• Provide libraries to access the data structures
• Broad support for many languages
• Create building blocks to form an ecosystem around it
• Implement adaptors for existing structures

15. Columnar Data

16. Previous Work
• CSV works really everywhere
• Slow, untyped and row-wise
• Parquet is gaining traction in all ecosystems
• one of the major features and interaction points of Arrow
• Still, this serializes data
• RAM-Copy: 10GB/s on a Laptop
• DataFrame implementations look similar but still are incompatible

17. Languages
• C++, C(glib), Python, Ruby, R, Matlab
• C#
• Go
• Java
• JavaScript
• Rust

18. There’s a social component
• It’s not only APIs you need to bring together
• Communities are also quite distinct
• Get them talking!

19. Shipped with batteries
• There is more than just data structures
• Batteries in Arrow
• Vectorized Parquet reader: C++, Rust, Java(WIP)
C++ also supports ORC
• Gandiva: LLVM-based expression kernels
• Plasma: Shared-memory object store
• DataFusion: Rust-based query engine
• Flight: RPC protocol built on top of gRPC with zero-copy optimizations

20. Ecosystem
• RAPIDS: Analytics on the GPU
• Dremio: Data platform
• Turbodbc: columnar ODBC access in C++/Python
• Spark: fast Python and R bridge
• fletcher (pandas): Use Arrow instead of NumPy as backing storage
• fletcher (FPGA): Use Arrow on FPGAs
• Many more … https://arrow.apache.org/powered_by/

21. Ecosystem
Kartothek:
• Heavily relies on Parquet adapter
• Uses Arrow’s type system which is more sophisticated than pandas’
• Using Arrow instead of building some components on their own allows
us to provide Kartothek access in other languages easily in the future

22. Does it work?

23. Does it work?
Everything is amazing on slides …

24. Does it work?
Everything is amazing on slides …
… so does this Arrow actually work?

25. Does it work?
Everything is amazing on slides …
… so does this Arrow actually work?
Let’s take a real example with:

26. Does it work?
Everything is amazing on slides …
… so does this Arrow actually work?
Let’s take a real example with:
• ERP System in Java with JDBC access (no non-Java client)

27. Does it work?
Everything is amazing on slides …
… so does this Arrow actually work?
Let’s take a real example with:
• ERP System in Java with JDBC access (no non-Java client)
• ETL and Data Cleaning in Python

28. Does it work?
Everything is amazing on slides …
… so does this Arrow actually work?
Let’s take a real example with:
• ERP System in Java with JDBC access (no non-Java client)
• ETL and Data Cleaning in Python
• Analysis in R

29. Does it work?

30. Does it work?

31. Does it work?

32. Does it work?
WIP

33. Get started easily?

34. Up Next
• Build more adaptors, e.g. Postgres
• Building blocks for query engines on top of Arrow
• Datasets
• Analytical kernels
• DataFrame implementations directly on top of Arrow

35. Thanks
Slides at https://twitter.com/xhochy
Question here!