The unit being shipped here is a standing object that holds state, not a statement that runs once and finishes. Databricks says an APPEND flow tracks new versus previously processed data in the source and manages the underlying serverless pipeline on its own [8]. That is the trade in one sentence: state that used to live in a checkpoint path or somebody's scheduler moves inside the platform, where it is not in a repo and not in a review.
Change data capture is where that matters. Databricks' own framing is that teams reach for MERGE INTO, then discover CDC records arriving out of order and needing extra logic to avoid incorrect results [9]. AUTO CDC replaces the hand-written version with declared keys, a sequencing specification, delete handling, and a choice of SCD Type 1 or Type 2 [10]. The difficulty did not disappear, it changed hands. Picking the sequencing column is the correctness decision in a CDC pipeline, and it is now made by whoever types the flow, in a surface built for querying, with no merge code left for a reviewer to argue with.
The named reference is making an argument about decomposition rather than about SQL. Adrien Marteau of bsport describes decoupling table loads from a single pipeline, processing third-party data independently, and getting better failure management and freshness out of it [11]. That benefit comes from breaking one pipeline into per-source units. It does not require those units to be written in the SQL editor; it does require them to be cheap enough to create that you make a lot of them.
Read the three patterns as a set and they are not equivalent. Append and REPLACE WHERE take away work that was mostly wasted: repeated insert logic with manual scheduling around it [4], and expensive full recomputes standing in for a partition or date-range overwrite [12]. Handing those to the platform costs a team very little judgment. CDC is the one where the hand-coded version carried judgment, and it is now the one-liner.
Two things the material does not settle. Only the append pattern is described as bringing its own automatically managed serverless pipeline; the CDC and REPLACE WHERE descriptions do not say the same [3]. And there is no pricing, compute, or availability detail for any of the three [2]. A flow declared in a query editor that then refreshes on a schedule indefinitely [7] is a recurring bill, and the announcement names no owner for it.
Worth keeping the scale honest: Databricks says thousands of SQL-first users already lean on Materialized Views and Streaming Tables [3], and that these patterns already existed as Lakeflow APIs [6]. The primitives are not the news. The authoring surface, and therefore the job title of the person who now operates them, is.
