hidi: --csdl-filter keeps every type, action and function, so a filtered CSDL still converts to the whole model #933

Summary

hidi transform --csdl … --csdl-filter <sets> only filters the EntityContainer. CsdlFilter.xslt keeps the matching EntitySet and Singleton elements and copies everything else unchanged: every entity, complex and enum type, every action and function, and every navigation property. On a large model, the result is still the whole model, and a client generated from it fails or is huge.

(Filed here because that's where hidi's source and CsdlFilter.xslt live; the NuGet package's project URL still points at microsoft/OpenAPI.NET.)

Repro

Dataverse Web API v9.2 $metadata (6.2 MB: 725 entity sets, 1,427 types, 16,005 navigation properties), Microsoft.OpenApi.Hidi 3.10.2, Kiota 1.38.0:

hidi transform --csdl metadata.xml --csdl-filter accounts,contacts -o out.yaml -f yaml -v 3.0 --co
kiota generate -l CSharp -c DataverseClient -n Issue.Client -d out.yaml -o out

The fixes below were applied one at a time, with the same filter and these conversion settings (--sp settings.json):

{ "OpenApiConvertSettings": { "EnablePagination": true, "EnableCount": true, "EnableNavigationPropertyPath": false, "EnableOperationPath": false, "EnableOperationImportPath": false } }

Proposal

Have --csdl-filter (or a companion option) keep only what the kept sets and singletons reach:

• the kept sets' entity types, their base types, and the entity types derived from them (so type casts still work);
• complex, enum and type-definition types reached through properties;
• bound actions and functions on kept types only, and unbound imports only when named;
• navigation properties whose target is kept. For navigation properties that leave the kept set, either drop them (opt-in) or keep their targets transitively.

The reachability trim we used is about 130 lines of Python over ElementTree (it doesn't follow derived types or bound operations yet). Happy to share it.

Related

• microsoft/OpenAPI.NET#2665 covers the same need at the OpenAPI level (CreateFilteredDocument). It concluded that the OpenAPI filter already trims as much as it can, because navigation properties link the models together. That makes the CSDL, before conversion, the place to cut.
• #65 asked how to filter a Dataverse CSDL and was answered with a hand-written XSLT. #389 is the same size problem on Business Central.
• Minor, from reading the XSLT: it matches set names with contains($entitySetOrSingleton, @Name), a substring test. With --csdl-filter accounts, a set named count or account would also be kept.