How are you preventing transaction couplings? For example, module A and B are called by C. C starts a transaction that wraps A and B. If you move B up as a network feature, you lose the transactionalty.
For a more concrete example, I recently built a service ("scraper") that scraped data and upserted a large tree of structured data to postgres in a transaction. Writes were only allowed from scraper, but "api" could SELECT data for reporting to the frontend "web" as much as it wanted. In the future, "api" might make be refactored to make internal HTTP request to "scraper," so they could have totally separate databases.