Artie deals with TOAST columns using 2 layers: 1. Reader layer: If REPLICA IDENTITY FULL is set on the table, Reader detects any TOASTed columns in the after-image and copies the real value from the before-image. So the message that goes to Kafka has the actual value, not the placeholder. 2. Transfer layer: Artie has a Transfer layer that maintains an in-memory row cache. If REPLICA IDENTITY FULL isn't set and the before-image wasn't in the WAL, Transfer pulls the previous value from the cache so the destination never sees the __debezium_unavailable_value placeholder. Artie Reader does emit the __debezium_unavailable_value string when it encounters a TOASTed column in the WAL because we built Transfer to be compatible with both Debezium pipelines and Artie Reader.

Schema drift: Artie never requires a schema registry. Schema changes are detected live from the source's mechanism (eg WAL for Postgres, DDL for MySQL) via Artie's Reader layer. If Reader sees a column that it doesn't recognize yet, it logs a warning and skips instead of crashing.

Let me know if you have any other questions!