Systemd Restart Policies

Choosing the Right Restart Policy

The Restart= directive controls when systemd automatically restarts a service. Each policy responds differently to various exit conditions.

Policy Comparison Table

Exit Condition	no	on-success	on-failure	on-abnormal	on-watchdog	on-abort	always
Clean exit (code 0)	-	Restart	-	-	-	-	Restart
Non-zero exit	-	-	Restart	-	-	-	Restart
Signal (SIGKILL, etc.)	-	-	Restart	Restart	-	Restart	Restart
Timeout	-	-	Restart	Restart	-	-	Restart
Watchdog timeout	-	-	Restart	Restart	Restart	-	Restart
Core dump (abort)	-	-	Restart	Restart	-	Restart	Restart

Recommended Policies by Service Type

Web Servers and APIs: Restart=on-failure

Restart=on-failure
RestartSec=5

Restarts on crashes but not on clean shutdown. This is the safe default for most services.

Critical Infrastructure: Restart=always

Restart=always
RestartSec=3

Always restarts regardless of exit reason. Use for services that must always be running (databases, message queues, monitoring agents).

Batch Jobs: Restart=no

Restart=no

The service runs once and stays stopped. Appropriate for oneshot tasks and migration scripts.

Services with Watchdog: Restart=on-watchdog

Restart=on-watchdog
WatchdogSec=30

Only restarts if the watchdog timer expires (service failed to send heartbeat). The service must call sd_notify("WATCHDOG=1") periodically.

RestartSec and Rate Limiting

RestartSec= adds a delay between restart attempts:

RestartSec=5     # Wait 5 seconds before restarting

Systemd also has built-in rate limiting. If a service restarts more than StartLimitBurst times within StartLimitIntervalSec, the service enters a failed state:

StartLimitIntervalSec=60
StartLimitBurst=5

This prevents a crash loop from consuming all system resources.

RestartPreventExitStatus

You can exclude specific exit codes from triggering a restart:

Restart=on-failure
RestartPreventExitStatus=1 6 SIGTERM

This restarts on failure except when the process exits with code 1, 6, or is killed by SIGTERM. Useful when certain exit codes indicate a configuration error that restarting won't fix.

Viewing Restart History

systemctl show myapp.service --property=NRestarts
journalctl -u myapp.service | grep "Started\|Stopped\|Failed"