Releases

v0.4.1

Introduce native Linux implementation using NETLINK_INET_DIAG. On Linux with the linux/inet_diag.h kernel header available, listener stats are collected via a single netlink round-trip instead of parsing /proc/net/tcp. Falls back to the pure-Ruby implementation automatically if the extension cannot be built.

Introduce Listener#fin_wait_count: connections in FIN_WAIT1 or FIN_WAIT2 state — server has sent FIN (initiated close) and is waiting for the peer to finish closing. Symmetric counterpart to close_wait_count.
Introduce Listener#time_wait_count: connections in TIME_WAIT state — both sides have closed; the kernel holds the socket for ~60s (2×MSL) to absorb delayed packets. High counts indicate fast connection churn.
Fix parse_address for IPv6 listeners: replace non-existent Addrinfo.parse with an explicit [ipv6addr].port pattern match.

Breaking Rename Listener fields: queue_size → queued_count, active_connections → active_count.
Introduce Listener#close_wait_count: number of accepted connections in CLOSE_WAIT state (peer has closed; application still processing).

Fixed queue_size under-reporting when multiple SO_REUSEPORT sockets share the same address — queue depths are now accumulated across all sockets rather than overwritten by the last one.
Linux Listener#active_connections for TCP no longer counts sockets that are still in the kernel accept queue (those remain in queue_size). Counts now match the usual “past accept()” meaning and align with tools such as Raindrops’ ListenStats#active.

Breaking IO::Metrics::Listener.capture returns an Array of Listener rows instead of a Hash keyed by address string.
Each Listener has address (Addrinfo for TCP or Unix), queue_size, and active_connections. Listener.zero sets address to nil. JSON uses Addrinfo#inspect_sockaddr for address, or null when absent.