The Catalyst Tree Manipulation Framework defines Expression abstraction for expression nodes in (the trees of) logical and physical query plans.
Spark SQL uses
Expression abstraction to represent standard and user-defined functions as well as subqueries in logical query plans. Every time you use one of them in a query it creates a new
When the query is executed,
Expressions are evaluated (i.e. requested to produce a value for an input InternalRow). There are two execution modes:
Be Careful With User-Defined Functions¶
There are many reasons why you should not write your own user-defined functions. First and foremost, they are a blackbox to Catalyst optimizer.
Speaking of memory usage, UDFs are written in a programming language like Scala, Java or Python that require an internal representation of data (InternalRow) to be fully deserialized and available as an object to the UDFs (that most of the time and for a reason know nothing about InternalRow and such). If it happens that two or more UDFs share computation (unless the UDFs are deterministic) they cannot share anything. Spark SQL cannot do much to optimize such queries.
There comes a thought that I'm still shaping in my head and haven't fully "dissected" yet.
Given that expressions (incl. UDFs) can be executed in code-generated execution mode that begs the question about possible performance improvements when an UDF uses
Expressions (as the "programming language"). I'm not really sure what the benefits could be yet, but gives some hope.
MonotonicallyIncreasingID expression is an example of "basic" expressions.