When using upstream PROXY protocol with tcp_proxy, a new connection pool is created in the cluster for each unique combination of downstream IP:port and upstream host.

For many typical use cases, where the downstream clients are not from a tightly constrained set of IP addresses, this results in a nearly-infinite number of connection pools, which wastes memory. Even if the clients are from a very small set of IPs, each could typically use 30,000 ephemeral ports, which could still result in a very large number of connection pools. The connection pools are never removed until the cluster is removed (via CDS, or process shutdown).

#13061 would have fixed this, but it was never completed/merged.