Linux watch command: rerun a command every few seconds

Run a command every N seconds

Running watch df -h clears your terminal, runs df -h every 2 seconds and fills the screen with the result. The top line shows a header such as Every 2.0s: df -h, followed by the hostname and the current time.

To change the wait between runs, pass -n (or --interval) with a number of seconds, or set the WATCH_INTERVAL environment variable for a new default. Fractions such as 0.5 work, anything below 0.1 is raised to 0.1, and anything above 2678400 is capped there.

watch df -h
watch -n 5 'ls -l /data'
watch -n 0.5 date

These check disk space at the default 2-second rate, list /data every 5 seconds and print the time twice a second. Press Ctrl+C to quit and return to your prompt.

The options worth knowing

These are the flags you will use most when monitoring jobs and system state:

procps-ng 4.0.4 on Debian 13 also has -q to exit when the output stops changing for a number of cycles, -r to skip reruns when the window is resized and -w to turn off line wrapping. Run watch --help to see what your version supports.

Quote pipes and shell syntax

watch passes your command string to sh -c. If you write a pipe, redirect or variable without quotes, your own shell handles it before watch starts, so it pipes the output of watch itself into the next command instead of rerunning the pipeline.

Wrap any command that uses |, > or $VAR in quotes. Single quotes keep your shell from expanding variables, so sh -c expands them fresh on every run, while double quotes expand them once before watch starts. If you don't need shell features, -x hands the command straight to exec, which saves you the extra quoting.

watch 'ps aux | grep [p]ython'
watch -n 1 'grep "^cpu MHz" /proc/cpuinfo | sort -nrk4'

Both keep the pipe inside single quotes so it runs on every update. The second example comes from the watch man page and sorts your CPU cores by clock speed once per second.

Everyday watch commands

When a training job, data pipeline or batch script is running, a few short commands cover most of what you check in a second terminal pane:

• nvidia-smi for GPU memory and utilization (see the nvidia-smi guide).
• free -h with -d, so memory and swap numbers light up as they move.
• df -h on a disk that a pipeline fills with large intermediate files.
• ls -lt | head -5 on an output directory to see the newest checkpoints arrive.
• kubectl get pods, though kubectl get also has its own -w (--watch) flag that streams changes.
• squeue --me on a Slurm cluster, with a long interval such as -n 60, because the Slurm docs ask you not to run squeue in loops.

watch -n 1 nvidia-smi
watch -d free -h
watch -n 2 'ls -lt /data/out | head -5'
watch -n 60 squeue --me

By default watch waits -n seconds after the previous run finishes. Pass -p to measure the interval from the start of the previous run instead.

Exit when the output changes

You don't have to stare at watch until something happens. -g (or --chgexit) makes watch exit with code 0 as soon as the output changes, so you can chain a follow-up command with &&.

watch -g -n 5 'ls /data/done' && echo "new file arrived"

This checks /data/done every 5 seconds and runs echo once a file appears or disappears. When the command you watch might fail, -e freezes the screen on a non-zero exit with command exit with a non-zero status, press a key to exit, and -b rings the terminal bell instead.

Where watch falls short

watch only draws the first screen of output, keeps no history of earlier runs, needs an open terminal and can only alert you inside that terminal, with a beep or by exiting. To check a long Python script or batch command from your phone, a live link beats a terminal polling loop.