Aims to help people upgrade to the latest version of Spark.
While we all wish we had a great test suite that covered all of the possible issues, that is not the case for all of our pipelines.
To support migrations for pipelines with incomplete test coverage, we have tooling to compare two runs of the same pipeline on different versions of Spark. Right now it requires that you specify the table to be compared, but (not yet started) for Iceberg tables we plan to provide a custom Iceberg library which automates this component.
Upgrading your code to a new version of Spark is perhaps not how most folks wish to spend there work day (let alone their after work day). Some parts of the migrations can be automated, and when combined with the upgrade validation described above can (hopefully) lead to reasonably confident automatic upgrades.
Spark SQL has some important changes between Spark 2.4 and 3.0 as well as some smaller changes in between later versions. (Spark SQL migration guide)[https://spark.apache.org/docs/3.3.0/sql-migration-guide.html] covers most of the expected required changes.
The SQL migration tool is built using, (SQLFluff)[https://sqlfluff.com/]. SQLFluff has a (Spark SQL dialect)[https://docs.sqlfluff.com/en/stable/dialects.html]
Out of the box SQLFluff lackes access to type information that is available when migrating Scala code, and the AST parser is not a 1:1 match with the underlying parser used by Spark SQL. A potential mitigation (if we end up needing type information) is integrating with Spark SQL to run an EXPLAIN on the input query and extract type information.
Some migration rules are too much work to fully automate so instead output warnings for users to manually verify.
We do not have an equivelent to "Scala Steward" for SQL files and SQL can target multiple backends. In most situations, the scheduler job type can be used to determine the engine.
The Python Upgrade tool is currently built using (bowler)[https://pybowler.io/], which is legacy-ish. This component has the least amount of development. See the next section for more information.
(Bowler)[https://pybowler.io/] can not parse Python 3.9+ only features. We should probably upgrade to (LibCST)[https://github.com/Instagram/LibCST] once it stabalizes.
LibCST (currently) has build issues that make upgrading to it questionable (e.g. the binder example fails to build libCST as of Oct 24th 2022).
Python migration tools do not have has much type information as Scala and are more likely to have conflicting APIs than in SQL (see pandas API overlap with Spark DataFrames). Logic to attempt to limit changes to only relevant Spark code may be problematic, so interactive fixes are more likely to be required.
The Scala upgrade tooling is built on top of ScalaFix and has access to (most) of the type information. Spark's Scala APIs are perhaps the fastest changing of three primary languages used with Spark.
While scalafix can be integrated with tools like Scala Steward (yay!), recompiling and publishing new artifacts is required to verify the changes. It is likely that dependencies will need to be manually upgraded.