The first flow graph uses aggregation. I don't see any issue of caching
or side-effects there as there are three separate input/output paths.
The second flow graph uses three separate "sub-steps" where each output
is appened to the current input document of the append step. From my
perspective, those are still three separate steps and so there are three
different results. If those sub-steps were sub-pipelines, you could
argue that they should produce the same result.
In the third example, the same step is shared amongst three append
steps. Here we have to decide on whether there is *caching* of results,
separate invocations, or user control of either behavior.
In all these examples, I don't see side-effects other than the question
of stable results.
Analogous to the document() function in XSLT, an issue for us is
whether flow graph 2 works the same when the output is labeled by a URI.
That is, is the document to URI mapping stable throughout the pipeline
execution?
--
--Alex Milowski
