e.g. Cobol files with dynamic definitions

Source stages accept source data, whether read from persistent storage or delivered in-memory, like a Kafka or MQSeries streaming source, and converts it to a 'virtual dataset': a table-based representation of the source data represented using one or more DataStage datatypes.  Sometimes this virtual dataset needs more processing before its data become useful.  For example, a source stage can generate a 'clob' of data in a single table cell which requires other stages (such as the pivot or hierarchical stages) to convert into a usable table of rows-and-columns which can be processed by other DataStage stages.

Put simply, MettleCI identifies your job's sources stages (those with no input links) and intercepts (and subsequently injects) data from/to all output links connected to that stage.  MettleCI is unconstrained by the contents of that link, and will use it whether it contains a conventional table of multiple rows and columns, or whether it contains a 'clob' in a single cell. 

Similarly, MettleCI identifies all target stages (those with no output links) and will inject the test data file described in your test specification into all output links connected to that stage, with not opinion on how many rows and columns that data comprises.