spark/databricks temporary views and uuid

[shadowinc]

Hi All,

We have a table which has an id column generated by uuid(). For ETL we use databricks/spark sql temporary views. we observed strange behavior between databricks sql temp view (create or replace temporary view) and spark sql temp view (df.createorreplacetempview()).

spark sql -  uuid() was evaluated every time and if joined by another table result was weird, uuid generated for 1 primary key column was asscoiated to another, somehow resulting in duplicates uuid()

df= select *, uuid() as id from source_table

df.createorreplacetempview(readData )

df= select * from readData join target_table on primary_key

df.createorreplacetempview(mergePrep  )

databricks SQL -  when using this and the same process, uuid() once generated were fixed and after joining also everything was fine.
readData = create or replace temp view readData  as select *, uuid() as id from source_table
mergePrep  = create or replace temp view mergePrep  as select * from readData join target_table on primary_key

Using databricks sql resolves my issue, however, I want to know how 2 approaches differ from each other while performing same operations. From my research I found that spark SQL df evaluates every time we use select, does that means even after creating temp view it evaluates (underlying nondeterministic functions like uuid), and same doesn't happen when using the databricks SQL method?

Appreciate your support on this. Point me to the right resources. Thanks

[Kaniz_Fatma · Accepted solution]

Hi @shadowinc,

1. Creation of Temporary Views:

   • Databricks SQL: When you create a temporary view using Databricks SQL, it’s scoped to the cluster an...¹.
   • Spark SQL: In contrast, Spark SQL’s createOrReplaceTempView method creates a temporary view scoped to the SparkSession. It gets dropped when the session closes².

2. Evaluation of Expressions:

   • Databricks SQL: Databricks SQL evaluates expressions (such as uuid()) once during view creation and retains the results. Subsequent queries against the view reuse these precomputed values.
   • Spark SQL: When using df.createOrReplaceTempView, expressions (including nondeterministic functions like uuid()) are re-evaluated each time you query the view. This dynamic evaluation can lead to different results, especially if the underlying data changes between queries.

3. Use Cases:

   • Databricks SQL: Ideal for scenarios where you want to create views from files (e.g., CSV, Parquet) or share views across notebooks within the same cluster.
   • Spark SQL: Useful when you need to work with DataFrames and want to export them as tables for data processing. It’s more flexible but may involve the recomputation of expressions.

Feel free to explore the provided resources for deeper insights! 😊🚀

[shadowinc]

Thanks, @Kaniz_Fatma  I suspected that, but could not find any links for confirming it.