Process information is collected non-atomically, so it can be possible that processes finish while metricbeat is collecting its metrics, provoking failures as the process is not found.
If something fails when collecting process information, whole events are lost, and errors are logged.

We could ignore the errors in some cases, reducing events loss.

We can also evaluate to use go-sysinfo instead of gosigar for processes info collection, at least on windows where it should be faster as mentioned in elastic/gosigar#110 (review)