I'm fine with this approach as well. Mostly I think it will just be helpful to say something about it in the spec or docs. This could mean prescribing specific behavior, or simply stating that the behavior is implementation-specific. Otherwise the next best thing for implementers to do is compare with existing drivers, which is what led to some of this confusion.
Any preference on leaving it open to backends or prescribing a specific behavior? On Tue, Apr 2, 2024 at 10:19 AM David Li <lidav...@apache.org> wrote: > I don't see why we should exclude them; I would also caution against > treating the driver behavior (especially only drivers in one language) as a > reference. > > On Tue, Apr 2, 2024, at 23:04, Joel Lubinitsky wrote: > > Hi, > > > > The ADBC spec does not currently define whether system > > catalogs/schemas/tables (e.g. information_schema.columns, sqlite_master, > > etc) should be included in the result of ConnectionGetObjects. > > > > A survey of existing driver implementations such as sqlite and postgresql > > indicates that the current convention is to exclude system objects, > > including only objects that have been defined by the user. Shall we > > formally add this to the spec? > > > > I think this would involve just updating a comment in adbc.h and any > > related documentation. The benefit would be better consistency for > drivers > > developed by vendors outside the arrow-adbc repo. > > > > For context, this came up in a PR [0] for the DuckDB ADBC implementation. > > > > Thanks, > > Joel > > > > [0]: https://github.com/duckdb/duckdb/pull/11446 >
