Restarting the snapshotter will cause nydusd to restart and exec pod error.

[ymygogogo]

1. systemctl stop containerd-nydus-grpc
2. systemctl start containerd-nydus-grpc
3. kubectl exec -it service-failure-nydus-57bf56f86c-s6dds -- sh

error: Internal error occurred: error executing command in container: failed to exec in container: failed to start exec "70ff624826aaa1a36a7f3913ae06a3cc00b97248944b39198af019a95755d514": OCI runtime exec failed: exec failed: unable to start container process: exec: "sh": executable file not found in $PATH: unknown

# /etc/systemd/system/containerd-nydus-grpc.service
[Unit]
Description=containerd-nydus-grpc
[Service]
Restart=on-failure
ExecStart=/usr/local/bin/containerd-nydus-grpc --nydusd-config /etc/containerd-nydus/nydusd-config.mirrors.json --log-to-stdout
[Install]
WantedBy=multi-user.target

[changweige]

By design, restarting nydus-snapshotter(containerd-nydus-grpc) should not affect running nydusd. Besides, even nydusd is not running there, the reported error should not indicate file not found. Perhaps, can you provide logs of nydus-snapshotter?

[BraveY]

Please provide the detailed steps to reproduce your issue. I can't reproduce the issue with just restarting nydus-snapshotter.

[BraveY]

As discussed in DingTalk, the issue stems from your systemd service configuration for nydus-snapshotter, which does not explicitly set KillMode=process. By default, systemd employs KillMode=control-group, which terminates all processes within the cgroup associated with the service. This includes both the parent process (nydus-snapshotter) and its child processes (e.g., nydusd).
To resolve this, explicitly configuring KillMode=process in the service unit file ensures that only the parent process is terminated, leaving child processes like nydusd unaffected and allowing them to persist as intended.