This commit adds a collection step in the Stop() task handler which will
retrieve the metrics available for this container at that time, and
store them until the next prometheus Collect() cycle.
This allows short-lived containers to be visible in prometheus, which
would otherwise be ignored (for example, running containerd-stress would
show something like 2 or 3 containers in the end, while now we can see
all of them). It also allows for more accurate collection when
long-running containers end (for example CPU usage could spike in the
last few seconds).
A simple case illustrating this with cpu usage would be:
ctr run -t --rm docker.io/library/alpine:latest mycontainer sh -c 'yes > /dev/null & sleep 3 && pkill yes'
Signed-off-by: Mathieu Pasquet <mathieu.pasquet@alterway.fr>
This adds an option for the cgroups monitor to include container metrics
in the prometheus output. We will have to use the plugin to emit oom
events via the events service but when the `no_prom` setting is set for
the plugin container metrics will not be included in the prom output.
Signed-off-by: Michael Crosby <crosbymichael@gmail.com>
This converts the oom metric to be a const metric so that deleted tasks
do not fill up the metric labels.
Signed-off-by: Michael Crosby <crosbymichael@gmail.com>
