A tool able to import existing pods into kueue.
The importer should run in a cluster having the Kueue CRDs defined and in which the kueue-controller-manager
is not running or has the pod
integration framework disabled. Check Kueue's installation guide and Run Plain Pods for details.
For an import to succeed, all the involved Kueue objects (LocalQueues, ClusterQueues and ResourceFlavors) need to be created in the cluster, the check stage of the importer will check this and enumerate the missing objects.
From kueue source root run:
go build -C cmd/importer/ -o $(pwd)/bin/importer
The command runs against the systems default kubectl configuration. Check the kubectl documentation to learn more about how to Configure Access to Multiple Clusters.
The importer will perform following checks:
- At least one
namespace
is provided. - For every Pod a mapping to a LocalQueue is available.
- The target LocalQueue exists.
- The LocalQueues involved in the import are using an existing ClusterQueue.
- The ClusterQueues involved have at least one ResourceGroup using an existing ResourceFlavor. This ResourceFlavor is used when the importer creates the admission for the created workloads.
The are two ways the mapping from a pod to a LocalQueue can be specified:
It's done by specifying a label name and any number of = a command line arguments eg. --queuelabel=src.lbl --queuemapping=src-val=user-queue,src-val2=user-queue2
.
It's done providing an yaml mapping file name as --queuemapping-file
argument, it's expected content being:
- match:
labels:
src.lbl: src-val
toLocalQueue: user-queue
- match:
priorityClassName: p-class
labels:
src.lbl: src-val2
src2.lbl: src2-val
toLocalQueue: user-queue2
- match:
labels:
src.lbl: src-val3
skip: true
- During the mapping, if the match rule has no
priorityClassName
thepriorityClassName
of the pod will be ignored, if more than onelabel: value
pairs are provided, all of them should match. - The rules are evaluated in order.
skip: true
can be used to ignore the pods matching a rule.
After which, if --dry-run=false
was specified, for each selected Pod the importer will:
- Update the Pod's Kueue related labels.
- Create a Workload associated with the Pod.
- Admit the Workload.
./bin/importer import -n ns1,ns2 --queuelabel=src.lbl --queuemapping=src-val=user-queue,src-val2=user-queue2 --dry-run=false
Will import all the pods in namespace ns1
or ns2
having the label src.lbl
set in LocalQueues user-queue
or user-queue2
depending on src.lbl
value.
With mapping file:
- match:
labels:
src.lbl: src-val
toLocalQueue: user-queue
- match:
priorityClassName: p-class
labels:
src.lbl: src-val2
src2.lbl: src2-val
toLocalQueue: user-queue2
./bin/importer import -n ns1,ns2 --queuemapping-file=<mapping-file-path> --dry-run=false
Will import all the pods in namespace ns1
or ns2
having the label src.lbl
set to src-val
in LocalQueue user-queue
regardless of their priorityClassName and those with src.lbl==src-val2
,src2.lbl==src2-val
and priorityClassName==p-class
in user-queue2
.