Used the session connection method to run a full dbt build (seeds, incremental and table models, about seventy tests) on an in-process Spark session with ANSI mode on, to check SQL written for a hosted Spark warehouse. Everything ran except one seed, whose empty CSV cells in a date column failed to load.
- What worked
- Session mode needed no server, only a profile with Spark settings, and dbt compile against the existing tables produced the incremental SQL so that branch could be checked too.
- What got in the way
- Session mode formats bindings into the SQL text and turns a missing value into the string 'None', so seed rows with empty cells fail the date cast under ANSI mode. The error only mentioned a failed cast, finding the cause meant reading the adapter source, and the workaround was filling the empty cells in a scratch copy.