An operation involving ownership, a credential, and a public identity needs a clear invariant across those records. I would identify which state must change atomically before splitting work into separate calls. The right boundary follows the business rule, not file organization.

The database accepting a migration is only one checkpoint. Existing readers, writers, indexes, and background jobs all need a compatible path. I would verify a representative deployed write and read after the schema transition rather than treating migration completion as release completion.

A slow collection endpoint can come from unnecessary row expansion, a poor join order, or a generic plan as well as a missing index. I start with a representative query, actual row counts, and the filter combinations callers use. What evidence do you keep with a proposed index?