While pgpool works well as an abstraction layer above PostgreSQL, its handling of excess client connection attempts is less than ideal. If the maximum number of clients per pool was twenty, for instance, any connections over twenty with the same login credentials and target database will simply wait indefinitely. Further, there is no concept of transaction-level connection reuse.

PgBouncer can allow prospective client connections to number in the thousands and still maintain high throughput. We can also tell it to reuse connections after any client completes a transaction so that clients do not have to disconnect between operations. Yet, it cannot balance connections across multiple PostgreSQL servers, and it certainly has no concept of primary server or replica. In this respect, it really is a bouncer, holding users at the door with minimal knowledge of what's inside the building.

Until there's a product that combines the best elements of these...