The process is executed as follows: dumb-init --single-child -> start.sh -> ovsdb-server.

When pod exits, k8s will send SIGTERM to dump-init, which will proxy it to start.sh but won't send TERM to ovsdb-server . start.sh will exit without sending SIGTERM to ovsdb-server either. In the end, k8s detects that the main process (start.sh) exited and sends SIGKILLs to remaining processes, incl. ovsdb-server.

We should give the process a change to exit cleanly. I think this can be done with something like:

diff --git a/templates/ovndbcluster/bin/cleanup.sh b/templates/ovndbcluster/bin/cleanup.sh
index bd3588e..fcd82df 100755
--- a/templates/ovndbcluster/bin/cleanup.sh
+++ b/templates/ovndbcluster/bin/cleanup.sh
@@ -44,3 +44,5 @@ if [[ "$(hostname)" != "{{ .SERVICE_NAME }}-0" ]]; then
     # now that we left, the database file is no longer valid
     rm -f /etc/ovn/ovn${DB_TYPE}_db.db
 fi
+
+/usr/share/ovn/scripts/ovn-ctl stop_nb_ovsdb

I don't think this is a critical problem, since the process will exit anyway. Maybe it affects RAFT leadership transition somewhat, since pod0 will not have a chance to let others know that it's exiting (I don't know if it communicates about it though.) But I don't have data to suggest it would e.g. corrupt db file. (I am told that invalid log entries in RAFT log are gracefully ignored by ovsdb-server.)