pd.{read/to}_sql is simple but not fast

Pandas provides convenience methods to read and write to databases using to_sql and read_sql. They provide great usability and a uniform interface for all databases that support an SQL Alchemy connection. Sadly, the layer of convenience also introduces a performance loss. Luckily, for a lot of databases, a performant access layer is available.

Transcript

1. pd.{read/to}_sql is simple but not fast Uwe Korn – QuantCo – November 2020
2. About me • Engineering at QuantCo • Apache {Arrow, Parquet} PMC • Turbodbc Maintainer • Other OSS stuff @xhochy @xhochy mail@uwekorn.com https://uwekorn.com
3. Our setting • We like tabular data • Thus we use pandas • We want large amounts of this data in pandas • The traditional storage for it is SQL databases • How do we get from one to another?
4. SQL • Very very brief intro: • „domain-specific language for accessing data held in a relational database management system“ • The one language in data systems that precedes all the Python, R, Julia, … we use as our „main“ language, also much wider user base • SELECT * FROM table
 INSERT INTO table
5. • Two main arguments: • sql: SQL query to be executed or a table name. • con: SQLAlchemy connectable, str, or sqlite3 connection
6. • Two main arguments: • name: Name of SQL table. • con: SQLAlchemy connectable, str, or sqlite3 connection
7. • Let’s look at the other nice bits („additional arguments“) • if_exists: „What should we do when the target already exists?“ • fail • replace • append
8. • index: „What should we with this one magical column?“ (bool) • index_label • chunksize: „Write less data at once“ • dtype: „What should we with this one magical column?“ (bool) • method: „Supply some magic insertion hook“ (callable)
9. SQLAlchemy • SQLAlchemy is a Python SQL toolkit and Object Relational Mapper (ORM) • We only use the toolkit part for: • Metadata about schema and tables (incl. creation) • Engine for connecting to various databases using a uniform interface
10. Under the bonnet
15. How does it work (read_sql)? • pandas.read_sql [1] calls SQLDatabase.read_query [2] • This then does
 • Depending on whether a chunksize was given, this fetches all or parts of the result [1] https://github.com/pandas-dev/pandas/blob/d9fff2792bf16178d4e450fe7384244e50635733/pandas/io/sql.py#L509-L516 [2] https://github.com/pandas-dev/pandas/blob/d9fff2792bf16178d4e450fe7384244e50635733/pandas/io/sql.py#L1243
16. How does it work (read_sql)? • Passes in the data into the from_records constructor • Optionally parses dates and sets an index
17. How does it work (to_sql)? • This is more tricky as we modify the database. • to_sql [1] may need to create the target • If not existing, it will call CREATE TABLE [2] • Afterwards, we INSERT [3] into the (new) table • The insertion step is where we convert from DataFrame back into records [4]
 
 [1] https://github.com/pandas-dev/pandas/blob/d9fff2792bf16178d4e450fe7384244e50635733/pandas/io/sql.py#L1320 [2] https://github.com/pandas-dev/pandas/blob/d9fff2792bf16178d4e450fe7384244e50635733/pandas/io/sql.py#L1383-L1393 [3] https://github.com/pandas-dev/pandas/blob/d9fff2792bf16178d4e450fe7384244e50635733/pandas/io/sql.py#L1398 [4] https://github.com/pandas-dev/pandas/blob/d9fff2792bf16178d4e450fe7384244e50635733/pandas/io/sql.py#L734-L747
18. Why is it slow? No benchmarks yet, theory first.
 
 
 
 
 
 
 
 
 

19. Why is it slow?
20. Thanks Slides will come after PyData Global Follow me on Twitter: @xhochy How to get fast?
21. ODBC • Open Database Connectivity (ODBC) is a standard API for accessing databases • Most databases provide an ODBC interface, some of them are efficient • Two popular Python libraries for that: • https://github.com/mkleehammer/pyodbc • https://github.com/blue-yonder/turbodbc
22. ODBC Turbodbc has support for Apache Arrow: https://arrow.apache.org/ blog/2017/06/16/turbodbc-arrow/
23. ODBC • With turbodbc + Arrow we get the following performance improvements: • 3-4x for MS SQL, see https://youtu.be/B-uj8EDcjLY?t=1208 • 3-4x speedup for Exasol, see https://youtu.be/B-uj8EDcjLY?t=1390
24. Snowflake • Turbodbc is a solution that retrofits performance • Snowflake drivers already come with built-in speed • Default response is JSON-based, BUT: • The database server can answer directly with Arrow • Client only needs the Arrow->pandas conversion (lightning fast⚡) • Up to 10x faster, see https://www.snowflake.com/blog/fetching- query-results-from-snowflake-just-got-a-lot-faster-with-apache- arrow/
25. JDBC • Blogged about this at: https://uwekorn.com/2019/11/17/fast-jdbc- access-in-python-using-pyarrow-jvm.html • Not yet so convenient and read-only • First, you need all your Java dependencies incl arrow-jdbc in your classpath • Start JVM and load the driver, setup Arrow Java
26. JDBC • Then: • Fetch result using the Arrow Java JDBC adapter • Use pyarrow.jvm to get a Python reference to the JVM memory • Convert to pandas 136x speedup!
27. Postgres Not yet opensourced but this is how it works:
28. How do we get this into pandas.read_sql?
29. API troubles • pandas’ simple API: 
 • turbodbc

30. API troubles • pandas’ simple API: 
 • Snowflake

31. API troubles • pandas’ simple API: 
 • pyarrow.jvm + JDBC

32. Building a better API • We want to use pandas’ simple API but with the nice performance benefits • One idea: Dispatching based on the connection class
 • User doesn’t need to learn a new API • Performance improvements come via optional packages

33. Building a better API Alternative idea:
34. Building a better API Discussion in https://github.com/pandas-dev/pandas/issues/36893
35. Thanks Follow me on Twitter: @xhochy