[metricbeat] check for a valid limit before we process a memory event

[fearful-symmetry]

This fixes #11283

It appears that we can get a stat object back when a container is in some odd in between state where it's not actually started. During this time, all the memory values we're using are zero, and we get NaNs from the division operations:

"memory":common.MapStr{"limit":0x0, "rss":common.MapStr{"total":0x0, "pct":NaN}, "usage":common.MapStr{"pct":NaN, "max":0x0, "total":0x0}, "fail":common.MapStr{"count":0x0}}}

The docker source states that Limit will only be 0 if a container isn't started, so we use that to check if we have a valid event. We can also process the event and just wrap the division in if blocks, but I liked the idea of skipping the event if we have zero'ed out data.

[odacremolbap]

+1
And awesome adding comments to existing funcs, thanks for that.
I guess that it is not a real race, docker exposes the endpoints as soon as the container object is created, and we are polling before there is any data there.

It would be probably very expensive to GET for the container status at every request, so checking for the value that breaks our code sounds right to me.

Not needed, but a bit more awesomeness and love if setting a proper variable name to
https://github.com/elastic/beats/blob/0d9bdcf21f41aa41598a60a8da808a245be14e01/metricbeat/module/docker/memory/helper.go#L43

something like

	for _, containerStats := range containersStats {

[fearful-symmetry]

@odacremolbap Yah, I kinda gave those variable name the side-eye earlier too.

[kaiyan-sheng]

a space here would be great :-) "// MemoryData..." and same in line 38

[ruflin]

Can you add a changelog entry? Also added needs_backport as I think we should fix this too in older versions.

[fearful-symmetry]

Looks like coredns is failing here now too:
Exception: Timeout while waiting for healthy docker-compose services: coredns

[ruflin]

Let's try to rebase.

[jsoriano]

It seems it was already backported, I am removing the needs_backport label.