On RELEASE_IP, the releasing node receives CTDB_CONTROL_TCP_CLIENT_DISCONNECTED, which causes the TCP connection to be removed from the tracking list on all nodes via CTDB_CONTROL_TCP_REMOVE. This means that the takeover node does not know about the TCP connection, so does not send tickle ACKs to the client. This was noticed after (finally) running CTDB cluster tests under autocluster for the first time in a long time and finding that ctdb/tests/CLUSTER/complex/32_cifs_tickle.sh was failing. Avoiding removal of the TCP connection during RELEASE_IP in ctdb_control_tcp_client_disconnected() fixes the issue. Sorry, I missed this when reviewing c6602b686b4e50d93272667ef86d3904181fb1ab in 2023. I will now regularly test via autocluster to ensure we don't see similar regressions.