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
Since a good deal of netdata's value is in keeping its' footprint as small as possible, and instances are stateless, adding default limits for child pods seem to make sense. I would include parent and k8s-state too, but AFAIK those are relative to the size/complexity of the cluster and therefore hard to predict.
Setting resource requests is just good practice. Setting limits prevent runaway processes.
Based on the docs I suggest defaulting the child to
I would include parent and k8s-state too, but AFAIK those are relative to the size/complexity of the cluster and therefore hard to predict.
Hi, @ohthehugemanatee. I think we have the same situation with the children nodes - it is hard to predict, depending on the number of pods per node and the workload (e.g. a lot of cronjobs (many short-lived pods/containers)).
Setting resource requests is just good practice
You are 100% right. As it is now - it is up to the user: to install, check resource usage and set the limits. The default values (not only the limits) are not production ready.
default values (not only the limits) are not production ready.
OK so this is the first problem to solve. From my understanding in a k8s situation the children should be very lightweight; it's the parent that really gets big. So a start would be to approach this only for the children.
What about using the calculation from docs as a starting point? Add the ephemerality and tiers-related config values to the helm chart values, so we can use helm arithmetic to arrive at a high estimate default value. In the output after install, add a line to say "we set memory limits based on an estimate of your child node RAM requirements. There is almost certainly room to reduce these limits, see https://learn.netdata.cloud/... for more information.
Even better would be to request feedback on a GH issue if the defaults suck, but you might get that regardless. 😃
Since a good deal of netdata's value is in keeping its' footprint as small as possible, and instances are stateless, adding default limits for child pods seem to make sense. I would include parent and k8s-state too, but AFAIK those are relative to the size/complexity of the cluster and therefore hard to predict.
Setting resource requests is just good practice. Setting limits prevent runaway processes.
Based on the docs I suggest defaulting the child to
The text was updated successfully, but these errors were encountered: