Introduce open_fds saturation point for process_exporter
AFAICT we have a gap in our file descriptor monitoring. While we do have an open_fds saturation point, this looks only at process_open_fds / process_max_fds. It applies only to in-process exporters that actually are able to self-report these metrics.
We have several services that are sensitive to open fd limits:
- gitaly
- haproxy
- redis
For gitaly we are able to rely on the in-process metrics. For haproxy and redis however, we rely on process_exporter.
process_exporter exposes a metric that does what we want: namedprocess_namegroup_worst_fd_ratio (source).
We are not currently monitoring this metric in saturation points and subsequently capacity planning. We can see that several of our redises actually have fairly high values here:
This was discovered in light of investigating #2754 (closed).