Single-region database with no caching of balances at all, since any cache layer is fundamentally incompatible with the requirement that balance values be exactly current.

Replicated database with an aggressive write-back cache for balances, because faster writes will improve the user experience even if the cache occasionally lags behind the database.

Replicated database with a long TTL cache on balance values, because long TTLs maximize the cache hit rate and reduce overall energy consumption across the system.

Single-region database with write-through caching for balance values, and routing of all balance reads to the primary database itself rather than to any read replica that might lag behind it.