filterOutSchedulableByPacking is an alternative to the older
filterOutSchedulable. filterOutSchedulableByPacking sorts pods in
unschedulableCandidates by priority and filters out pods that can be
scheduled on free capacity on existing nodes. It uses a basic packing
approach to do this. Pods with nominatedNodeName set are always
filtered out.
filterOutSchedulableByPacking is set to be used by default, but, this
can be toggled off by setting filter-out-schedulable-pods-uses-packing
flag to false, which would then activate the older and more lenient
filterOutSchedulable(now called filterOutSchedulableSimple).
Added test cases for both methods.
Replace the simple boolean ScaledUp property of ScaleUpStatus with a more
comprehensive ScaleUpResult. Add more possible values to ScaleDownResult.
Refactor the processors execution so that they are always executed every
iteration, even if RunOnce exits earlier.
- This is intended to address the issue described in https://github.com/kubernetes/autoscaler/issues/923
- the delay is configurable via a CLI option
- in production (on AWS) we set this to a value of 2m
- the delay could possibly be set as low as 30s and still be effective depending on your workload and environment
- the default of 0 for the CLI option results in no change to the CA's behavior from defaults.
Change-Id: I7e3f36bb48641faaf8a392cca01a12b07fb0ee35
No change in actual logic being executed. Added a new
NodeGroupListProcessor interface to encapsulate the existing logic.
Moved PodListProcessor and refactor how it's passed around
to make it consistent and easy to add similar interfaces.
AutoscalingContext is basically a configuration and few static helpers
and API handles.
ClusterStateRegistry is state and thus moved to other state-keeping
objects.
The CleanUp method is instead called directly from the implementation,
when required.
Test updated in a quick way since the mock we're using does not support
AtLeast(1) - thus Times(2).
Nodes with GPU are expensive and it's likely a bunch of pods
using them will be created in a batch. In this case we can
wait a bit for all pods to be created to make more efficient
scale-up decision.
Jacek Kaniuk
Vivek Bagade
Łukasz Osipiuk
Kubernetes Prow Robot
Maciej Pytel
Thomas Hartland
Aleksandra Malinowska
Jakub Tużnik
Steve Scaffidi
Krzysztof Jastrzebski
Beata Skiba
Karol Gołąb
anniedy