inotify support for kubernetes_logs #20541
Labels
source: kubernetes_logs
Anything `kubernetes_logs` source related
type: feature
A value-adding code addition that introduce new functionality.
A note for the community
Use Cases
Some pods may succeed (enter the
Succeeded
state) very quickly. The lifecycle of the log folder is presently tied to the pod lifecycle itself, and the duration during which it is tracked is not straightforward to control (it is actually unclear to what extent that can be controlled at all).Attempted Solutions
1.1-2.1x cooldown
to restore perfect delivery. That is the best solution, but requires workload-author involvement.glob_minimum_cooldown_ms
from 60s to 5s polling time lowers the amount of impacted workloads at the cost of vastly increased CPU usage. There must be side-effects to the node itself which has to service the extra filesystem syscalls as well.Proposal
I can't do much outside of expressing my strong support for
inotify
support for the kubernetes source. That may or may not be implemented as part of the file source. It could be an entirely new source on which to base the kubernetes source.In some of the referenced issues, an argument for platform-independence was made. While the
file
source certainly has to be platform independent, why should thekubernetes
one be? I claim no expertise on the matters of Kubernetes on Windows, but a quick look at the Kubernetes | Windows User Guide | Capturing logs from workloads indicates that the prefered way is not to use files at all. Because of that, it seems to me that thekubernetes
source does not support Windows adequately today - and as such, the platform independence argument is moot. Outside of Linux, Kubernetes only runs on Windows as far as I'm aware (k8s/BSD is hardly a thing).One should not have to choose between efficiency and deliverability, which is the tradeoff we have to make here. Let's have both with
inotify
!References
default_glob_minimum_cooldown_ms
too high, does not document its existence 聽#6771 is a trace of a past changeglob_minimum_cooldown_ms
default of 5 seconds correct?聽#7840 is a long discussion of the tradeoffs for different values of the polling timerVersion
vector 0.36.0 (x86_64-unknown-linux-gnu)
The text was updated successfully, but these errors were encountered: