The PostgreSQL connector (and others) doesn't directly deserialize from the DbDataReader; there's a temporary Dictionary<string, object> "storage model", which is read from the DbDataReader and then the CLR type is populated from that. This adds an additional unnecessary intermediate conversion, memory allocations for the temporary dictionary, and various dictionary lookups. We should change the relevant connectors to read and write directly instead. The dictionary also forces boxing of value types to occur (#11183).

Note: this will prevent the user of custom mappers in their current form, since they're designed around the idea of converting one object to another; but when dealing with ADO.NET, database results aren't an object - they're exposed via DbDataReader which must be exposed. So a PostgreSQL custom mapper would need to accept a DbDataReader and implement the logic for reading properties out of it directly (and for the other side - writing - it would need to accept or return a DbCommand, on which it would populate DbParameters).