Issue
In some environments running with a Network Load Balancer in front of Redpanda, it was observed there were a higher number of http connections to Redpanda (HTTP Proxy/schema registry etc) than expected per metric vectorized_httpd_connections_current
Detail
Investigations showed that the http connections to Redpanda were being silently dropped by some Network Load balancers but never being cleaned up on the Redpanda side
A fix was created for Redpanda to implement TCP keepalive on the Redpanda HTTP API's
Resolution
The keep-alive fix was implemented via PR: http: Add TCP keepalive to http APIs #28838
which is available in Redpanda Release 25.3.2 and higher