My team is using PHP-FPM, configured statically with a max of 32 processes, in a Kubernetes cluster. We're autoscaling based on PHP-FPM metrics, and there are many replicas. I'm frequently (i.e. on all replicas) seeing the PHP-FPM status show queued connections, even though the max active processes is below total processes. For example:

pool:                 www
process manager:      static
start time:           06/Feb/2025:17:37:07 +0000
start since:          80556
accepted conn:        265772
listen queue:         0
max listen queue:     2
listen queue len:     511
idle processes:       27
active processes:     5
total processes:      32
max active processes: 21
max children reached: 0
slow requests:        0

Note that the max active processes is 21 of the 32 available, but that the listen queue has reached 2. Why are connections being queued when there are idle processes?