You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The goal of this PR is to have a cAdvisor metric which (as accurately as possible)
describes the amount of container memory which is not evictable by the kernel.
This new metric can be used to accurately graph and alert on
container memory usage regardless of its evictable memory usage patterns (e.g large active page cache).
working_set_bytes today does not always align with non-evictable memory.
For example, two containers in a pod sharing files in an emptyDir will increase total_active_file cache
as one container writes and another container reads over time, dramatically increasing working_set_bytes.
Under increasing non-evictable memory demands from the file owning process, total_active_file will decrease, and working_set_bytes's value will hover around ~90% of the cgroup memory limit.
This makes alerting difficult, as working_set_bytes does not accurately detail that the pod has evictable active page cache that the kernel is slowly draining.
In other words, total_active_file memory can be evicted by the kernel, but is included in working_set_bytes.
Alternatively to a new metric, working_set_bytes could be updated to represent non evictable memory and exclude total_active_file (along with any other evictable fields).
0 commit comments