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.
When a machine with GPU becomes ready it can take
up to 15 minutes before it reports that GPU is allocatable.
This can cause Cluster Autoscaler to trigger a second
unnecessary scale up.
The workaround sets allocatable to capacity for GPU so that
a node that waits for GPUs to become ready to use will be
considered as a place where pods requesting GPUs can be
scheduled.
This can be used to dynamically update cloud provider
config (in particular list of managed NodeGroups and their
min/max constraints).
Add GKE implementation.
Currently we track if those nodes can be removed and only
skip them at the execution step. Since checking if node is
unneeded is pretty expensive it's better to filter them out
early.
This flag was true in default setups for every platform,
we haven't heard about any user changing it to false and
after removing check on PodScheduled condition setting it
to false would basically break CA.
By fixing CA not to reset `StaticAutoscaler` state before each iteration so that it remembers last scale-up/down time which is used to throttle scale-down, which is causing the issue.
Automatic merge from submit-queue
Cluster-autoscaler: include PodDisruptionBudget in drain - part 1/2
In part 1 or 2 we skip nodes that have a pod with 0 poddisruptionallowed. Part 2/2 will delete pods using evict.
cc: @jszczepkowski @MaciekPytel @davidopp @fgrzadkowski
StaticAutoscaler.kubeClient was uninitialized,
leading to segfaults when trying to use it. It was
also a duplicate since the client is already available
through AutoscalingContext.
Adds a new optional flag named `configmap` to specify the name of a configmap containing node group specs.
The configmap is polled every `scan-interval` seconds to reconfigure cluster-autoscaler dynamically at runtime.
Example usage:
```
./cluster-autoscaler --v=4 --cloud-provider=aws --skip-nodes-with-local-storage=false --logtostderr --leader-elect=false --configmap=cluster-autoscaler --logtostderr
```
The configmap would look like:
```yaml
kind: ConfigMap
apiVersion: v1
metadata:
name: cluster-autoscaler
namespace: kube-system
data:
settings: |-
{
"nodeGroups": [
{
"minSize": 1,
"maxSize": 2,
"name": "kubeawstest-nodepool1-AutoScaleWorker-1VWD4GAVG35L5"
}
]
}
```
Other notes:
* Make namespace defaults to "kube-system"
according to https://github.com/kubernetes/contrib/pull/2226#discussion_r94144267
* Trigger a full-recreate on a configuration change
according to https://github.com/kubernetes/contrib/pull/2226#issuecomment-269617410
* Introduced `autoscaler/` and moved all the dynamic/recreatable-at-runtime parts of autoscaler into there (Update: the package is now named `core` according to https://github.com/kubernetes/contrib/pull/2226#issuecomment-273071663)
* Extracted the core of CA(=`func Run()` in `main.go`) into `Autoscaler`
* `DynamicAutoscaler` is a wrapper around `Autoscaler` which achieves reconfiguration of CA by recreating an `Autoscaler` instance on a configmap change.
* Moved `scale_down*.go`, `scale_up*.go` and `utils*.go` into the `autoscaler` package accordingly because they seemed to be meant to be collocated in the same package as the core of CA (which is now implemented as `Autoscaler`)
* Moved the `createEventRecorder` func from the `main` package to the `utils/kubernetes` package to make it importable from both `main` and `autoscaler`
Aleksandra Malinowska
Krzysztof Jastrzebski
Beata Skiba
Maciej Pytel
Karol Gołąb
anniedy
Marcin Wielgus
Edward Tsang
Matt Terry
Ivan Towlson
Yusuke Kuoka
Kubernetes Submit Queue