UDFs & Table Functions

Stories about user-defined functions — their cost, their limitations, and more powerful alternatives.

Stories

• The UDF Tax: Why User-Defined Functions Are a Black Box to the Optimizer — Catalyst opacity, Python UDF overhead, deserialization cost, when to use built-ins instead
• One Row In, Many Rows Out: The Story of User-Defined Table Functions — explode and built-in generators, custom Python UDTFs, LATERAL VIEW, table arguments

Related stories

• Two Runtimes, One Job: How PySpark Bridges Python and the JVM — Python UDFs pay the full Py4J bridge overhead described here for every row
• Pandas UDFs: How Arrow Makes Python Functions Fast Enough for Spark — the Arrow-based alternative to row-at-a-time Python UDFs
• From SQL to a Running Plan: The Catalyst Story — why UDFs are opaque to Catalyst’s optimization rules
• Expressions All the Way Down: How Spark Represents and Evaluates Computations — built-in functions are expression tree nodes; UDFs are not, which is why they can’t be optimized