Allow StructureSpecific in XML writer

[brockfanning]

If I'm understanding the DataMessage section in the XML writer code it will always produce "Generic" output. Would it be possible to allow "StructureSpecific" as an alternative?

If so, and you can summarize what needs to happen, I would be happy to submit a pull request. I tried locally just by changing "Generic" to "StructureSpecific" in the line linked above, but that does not alter the inner structure. For example, the output still has "SeriesKey" elements, which to my knowledge is not in StructureSpecific.

[khaeru]

Certainly possible and a desirable enhancement, but non-trivial.

The summary: the "inner structure" is generated by the rest of the code in the function you linked to. Every call to writer.recurse() within that function results in a call to a different method in the same file. For instance:

sdmx/sdmx/writer/xml.py

Lines 130 to 131 in 3e4cbc1

	# Add data
	elem.append(writer.recurse(ds))

In this case, ds is an instance of model.DataSet. So the writer finds another method that accepts this type of object:

sdmx/sdmx/writer/xml.py

Lines 541 to 542 in 3e4cbc1

	@writer
	def _ds(obj: model.DataSet):

…and calls that, which in turn recurses for the constituent classes like SeriesKey, Observation, etc. Each method returns lxml Element objects, that are assembled (e.g. elem.append()) into the complete XML document.

In order to output structure-specific SDMX-ML, every such method would need to be extended to (a) check whether generic or structure-specific output is to be produced and (b) if the latter, generate the XML elements and attributes that the standard requires for structure-specific data.

(The experiment you mention only changed the name of one particular generated XML element—because it did not alter any of the other code that generates the elements and attributes, the output does not change.)

A PR for this would also need appropriate extensions to the test suite.

[brockfanning]

@khaeru Thanks, the makes sense. Do you think the Generic and StructureSpecific are pretty much the entirety of the possible formats? Or should this be done in a way that allows for other formats in the future?

[khaeru]

Coding a standards-compliant implementation will require careful attention to the actual standards documents, which will answer that and other questions 🙂 The SDMX website seems to be down at this moment (?!), but here's an archived version: https://web.archive.org/web/20200912164322/https://sdmx.org/?page_id=5008 —§3A is the one you want to consult.

For the moment, if each of the affected functions in sdmx.writer.xml contains an if/else block with just 2 clauses, that's fine from my point of view. If it became a performance concern (cf. #49) or the code ends up being too complex, then we could try a pattern of using one set of functions for generic data but a distinct set of functions for structure-specific data. Let's cross that bridge if we come to it.