Abstract
Many of us have a love hate relationship with Apache Spark, especially when it comes to PySpark. Especially in Python, the UDF performance can be a dumpster fire. This fire comes from a glorious, toxic mix of data copies, serialization (pickle, json, etc.), and incompatibility with most accelerators (excluding NV Rapids).
This talk looks at the new work being done to allow us to bring the computation to the data rather than the data to the computation. The recently accepted Spark Project Improvement proposal adds basic, and safe, transpilation which can accelerate relatively simple Python UDFs by converting them into catalyst expressions; but safe only gets us so far. We'll look at how we can go beyond "safe" transpilation and use LLMs to dynamically convert large chunks of code for performance, how to pick the appropriate level of safety in doing so, and why this is all a little bit crazy and turned off by default (but the non-LLM transpilation should be safe. probably.)
Come for the performance, stay for the non-deterministic compilers with dynamic property testing.
