Update & Merge - scanning only required columns

[idosilverwater-vast]

requesting help regarding implementation of MERGE operation in a trino connector.

when executing UPDATE <table> SET col2 = 5 where col3 = 2
in a table with schema col1:int, ..., col5:int

the engine plans and executes a scan with all the columns (and rowid column)
although i'd expect only the rowid and col3 to be scanned (and if col3=2's pushdown is supported then only for rowid to be scanned).

I thought that maybe changing this behavior could be done through beginMerge but that is only called after applyProjection/applyPredicate. And in applyProjection I might not have enough information to correctly drop columns.

(I am using RowChangeParadigm.CHANGE_ONLY_UPDATED_COLUMNS
and returning in getMergeRowIdColumnHandle a column handle that doesn't depend on col1-5)

Is there a standard way of implementing such a thing?

thank you

[idosilverwater-vast]

As far as I understand (might be completely making up stuff) the optimizer cannot optimize out the unneeded columns because they are referenced in the merge row. And therefore the Trino optimizer cannot safely remove them.

In StatementAnalyzer.java, it seems that the mergeRowType includes all the dataColumnSchemas that are all the non hidden columns.

Maybe the mergeRow should be only the mergeCaseColumns updateCaseColumns and rowId (and columns needed for partition shuffling)

[idosilverwater-vast]

In QueryPlanner.java in public PlanNode plan(Update node)
that is the place in which the merge rowblock is built with all the cols (going by the comment - on purpose)

        // Build the merge rowblock, containing
        // All data columns in table order
        // The boolean present field, always TRUE for update
        // The tinyint operation number, always UPDATE_OPERATION_NUMBER for update
        // The integer merge case number, always 0 for update