I wasn't criticizing Postgres as much as I was a bit confused why I don't hear more people talking about this issue.
I think poolers confuse the issue because they solve the problem of allowing multiple processes to share the same pool, they don't increase throughput to the server.
The 5000 connection example was probably missing some context. I actually was using 10mb per connection as the overhead which creates a decently expensive server for not a lot of throughput.
By pipelining what I mean is allowing multiple requests to be in flight on the connection, each tagged with an ID that gets queued on the server. When a worker process gets around to processing it, it'll send back a response with the same ID. This increases driver complexity but also makes it so you need very few connections per client.
I think poolers confuse the issue because they solve the problem of allowing multiple processes to share the same pool, they don't increase throughput to the server.
The 5000 connection example was probably missing some context. I actually was using 10mb per connection as the overhead which creates a decently expensive server for not a lot of throughput.
By pipelining what I mean is allowing multiple requests to be in flight on the connection, each tagged with an ID that gets queued on the server. When a worker process gets around to processing it, it'll send back a response with the same ID. This increases driver complexity but also makes it so you need very few connections per client.