Add NetStat to parse /proc/net/stat/...

[AlexZzz]

This PR is a step to expose arp/ndisc/route cache stats from /proc/net/stat/. It was described in the issue prometheus/node_exporter#1626.

I decided to make it similar to lnstat tool, so it will parse all new files in /proc/net/stat in future.

[pgier]

Thanks for the contribution! I added a few in-line comments. I think the struct could be changed to better match the structure of the data. For example, each header name could be the key of a struct which points to a slice of the cpu values.

type NetStat struct {
    Filename string
    Stats map[string]uint64[]
}

Also, since this is a new feature, can you add a test for this?

[AlexZzz]

Thank you for the review!
I've changed data structure, also added a test.

[discordianfish]

LGTM!
Just not 100% sure about the name. Maybe we should just call it NetStats?

[AlexZzz]

I thought about the name, but there's /proc/net/netstat file already, and it is used by node-exporter netstat collector, so netstats name will add mess, IMO.

It could be named netcache, or something related to network-and-cache. But if there is a new file under /proc/net/stat and it is not about cache, it will be confusing again:)

[discordianfish]

Oh right, hrm yeah I guess lnstat is fine then - unless somebody else has a better idea. @SuperQ @pgier

[AlexZzz]

@pgier @SuperQ , any suggestions?

[SuperQ]

We still need to port over the /proc/net/netstat (and snmp/snmp6) parsing from the node_exporter. So we don't have a naming conflict to worry about. We do have a bit of a precedent for /proc/net/sockstat -> NetSockstat. I wish there was some documentation about this outside of the lnstat man page. I can't find anything in the usual kernel.org/doc.

Maybe slightly confusing, but how about we follow this pattern:

• /proc/net/netstat -> NetNetstat
• /proc/net/stat/* -> NetStat

[AlexZzz]

* `/proc/net/netstat` -> `NetNetstat`
* `/proc/net/stat/*` -> `NetStat`

I think this looks good, because it looks like a naming convention really: name consists of path. I'll rename Lnstat to NetStat then. And here we don't care about the name of the collector, right? I mean, we're leaving lnstat and netstat collectors names and they'll work with NetStat and NetNetstat structs, right?

[SuperQ]

If you're talking about the node_exporter collector name, Yes, we could call it --collector.lnstat to match the normal user utility. That seems right to me. What do you think @discordianfish?

[SuperQ]

When digging through some other issues/PRs, I discovered we already have some partial parsing of /proc/net/stat in #254.

How do we want to handle this? We have prometheus/node_exporter#1155 open to add handling of nf_conntrack to the node_exporter, but this would be duplicated by lnstat.

@pgier Should we use this generic parser instead of the conntrack parser?

[AlexZzz]

@pgier , @SuperQ , how is it going?
I would like to vote for this generic parser:)

[discordianfish]

Ok, let's merge this. We can make net_conntrackstat.go use the parser here.

[AlexZzz]

@discordianfish , btw, I also have a wip pr with more generic metrics exposer which will expose conntrack stats if they appear at /proc/net/stat: prometheus/node_exporter#1771

[discordianfish]

What should we do with this now that prometheus/node_exporter#1155 has been merged?

[AlexZzz]

I suggest, prometheus/node_exporter#1155 might be reverted in favor of this one 😂

IMHO, it's better to have generic parser for any metric and any file, instead of hard-code here. Kernel API shouldn't be changed and files format should always be the same. Any new metric will be exposed with no human interaction with generic parser.

[discordianfish]

@AlexZzz I agree, lets rebase this and merge it, then create an issue to track using this to replace the conntrack stats in node-exporter.

[AlexZzz]

@discordianfish , rebased.

[discordianfish]

cc/ @SuperQ

[SuperQ]

LGTM