Open maxime1907 opened 1 month ago
you are not running in docker, it's the --path.procfs
correct?
Or is there any error in the log?
maybe it's dial cost too many time https://github.com/prometheus/node_exporter/blob/b9d0932179a0c5b3a8863f3d6cdafe8584cedc8e/collector/arp_linux.go#L79-L83
The pprof SVG appears to be truncated, and only shows the collection taking 1.84s (which is admittedly still kinda slow). I can't think of a reason off the top of my head for the rtnetlink variant of the arp collector to be slow. If anything, it should be faster than parsing /proc/net/arp
, especially on systems with a large number of ARP entries.
Can you please attach a complete pprof, so that it reveals where that ~1.8s is being consumed by the arpCollector?
The gaps in your Grafana graphs also suggest that you're getting the occasional scrape timeout. Is anything unusual logged when that happens?
Host operating system: output of
uname -a
node_exporter version: output of
node_exporter --version
node_exporter command line flags
node_exporter log output
Are you running node_exporter in Docker?
No, we use containerd directly
What did you do that produced an error?
Run node exporter which uses by default the collector collector.arp.netlink
What did you expect to see?
Normal scrape times like 300ms
What did you see instead?
Scrape times of more than 20 seconds
Here is a graph of before and after the deactivation of the arp netlink collector with this flag:
Here is the output of pprof with: