prometheus-dirsize-exporter

Export directory size metrics efficiently.

Why?

When providing multi-user interactive computing services (with a HPC cluster or with JupyterHub), it's very helpful to know the home directory sizes of each user over time. However, as NFS is often used, running du constantly takes a long time, uses too many IOPS that we may not have many of, and is plain inefficient.

This project provides a way to keep track of directory sizes with a budgeted amount of IOPS. You can ask it to take however much time it needs but not use more than 100 IOPS, and it will do that. We do not necessarily need very up to date directory size metrics, so this is ok.

Limitations

• As directory contents might change in the course of a single run as we wait for budgets to become available, information about a directory may not be exactly correct immediately.
• Because we do not spread the IOPS through time, the IO usage is 'spiky' - all IOPS get done at the beginning of a second, and then it goes silent. If you have a big IOPS budget, this can cause performance degradation.

Installation

Install the package from PyPI:

pip install prometheus-dirsize-exporter

Running

You can start the process from the commandline:

dirsize-exporter <path-to-parent-directory> <iops-budget> <wait-time-in-minutes>

For example, to export metrics about directories under /home, using no more than 200 IO operations per second, updating every 60minutes, you would run:

dirsize-exporter /home 200 60

You can check out the metrics by hitting http://localhost:8000. The port can be controlled via a --port argument.