hi, for delta_lake's merge_into grammar, the document says we need to specify the source, it can just be a normal table, it can be an external table or other things like data interface?

The source can be anything that you can turn into a DataFrame (when using the Scala or Python API) or anything that you can turn into a SQL query (when using the SQL API). It could even be the result of joining a 100 tables.

so the source will be a complete component forever，never split it？

I’m not sure if I understand the question. Could you rephrase?

I think if the source is a query, you will treat the whole query as a complete source or try to use join and the pass it to the optimizer?

Your question is still not clear. Internally merge does two joins between the source and the target table, and those are indeed passed through the optimizer. The source is a DataFrame and it corresponding lineage and properties.

why not use only once join but use two joins? Ae there some special reasons?