In this case, we're paging based on a table we JOIN, so we can just include the column we care about in the SELECT and then read the raw row value.

Pretty sure nothing in the UI actually uses this today: I don't think you can execute an interface query by name with a limit and page through it? Maybe in a typeahead? I'll go double check that.

This one isn't too bad:

• Feed stories have no "ID", use "chronologicalKey" (basically a 64-bit ID) instead.
• They return a raw row instead of an object since they're super old. This is fine, just slightly odd.

Tested this by paging through feed locally, seemed to work properly.

Maniphest is the biggest mess by far.

You can "Group By: Project" which allows each result to appear more than once in the result set. If task T123 is in projects "A" and "B", you'll get back two copies of T123 in the results.

We should maybe just remove this (it predates workboards by several years, although they aren't really the same thing), but I was able to keep it working without more than ~50 lines of gross disgusting mess.

(This is redundant, we automatically load all the columns from the primary table. Query ends up looking like SELECT r.*, r.*, ...).

This is only used by edge.search.

Edges have no IDs, so we're faking it. This isn't too awful. Some day, maybe they should have IDs.

This could have been removed earlier, I got rid of the only callsite near the beginning of this change sequence.

I changed the modular datasource limit from 100 to 3, then browsed interfaces in the "Bind Interface" UI from a Service. Worked fine.