The problem is that even a single user matrix server can be very resource intensive if that user joins big rooms with thousands of users spread over thousands of servers. Synapse is very database heavy, so the parallelism in Postgres helps a lot - plus some of the hot DB paths have special cased queries for Postgres to use some of its more obscure features that Sqlite lacks. Finally, we don’t dogfood or optimise Synapse with Sqlite, so there’s a risk of perf regressions.