The architecture of varnish is based on a management process supervising its child. The child does all the actual work of handling requests, caching, and so forth.
In some unfortunate cases, the child process can crash. When that happens, the management process detects the event by failing to communicate with its child, and tries to kill it.
Due to CAP_KILL not being in the Capability Bounding Set defined in varnish-frontend systemd unit, sending such signal currently fails:
Jan 09 20:59:52 cp3062 varnishd: Failed to kill child with PID 112629: Operation not permitted