A database is an implementation boundary
If Product joins tables owned by Seller, a Seller schema change becomes a Product deployment concern. Local database access may be fast, but the organizational and release coupling remains.
Direct writes are even more dangerous because they bypass the rules owned by Seller. Keep the service responsible for its own invariants.
Copy the view, not the ownership
Product often needs only a small subset of seller information to answer a query. A local read model can hold that subset without becoming the authoritative seller record.
Define which fields Product needs and how changes reach the projection. Avoid copying the whole Seller schema simply to reproduce remote joins locally.
Choose consistency deliberately
A local projection may lag behind its source. Decide which decisions can tolerate that delay and which need an authoritative check during a business operation.
Document update, deletion and recovery behavior. Removing a database dependency is useful only if the replacement contract is explicit.
Further reading
Primary documentation & specification