README
¶
The allocator code exists in three variants which get picked depending on which features are enabled:
- stable: for a "GA only" feature configuration, minimal changes going forward
- incubating: the default implementation, adds support for beta features
- experimental: under active development, including alpha features
This structure serves as a safety net because experimental changes cannot break more stable Kubernetes configurations, something that happened already once despite careful reviews.
The goal is to rotate the implementations wholesale instead of copying individual code chunks, i.e. at some point "incubating" replaces "stable", "experimental" replaces "incubating", and "experimental" becomes a copy of "incubating" until new changes get added to it again.
Ideally changes should be limited to "experimental", but sometimes changes have to be applied the same way across different variants, for example bug fixes or changes to the package API.
Tests are shared between all implementations, with test cases applied depending
on what features they require. Further testing is covered by
test/integration/scheduler_perf. When promoting implementations, the selection
of implementations which support certain features there needs to be
updated. For example []string{"experimental"} in TestSchedulerPerf of
test/integration/scheduler_perf/dra/consumablecapacity/consumablecapacity_test.go
will eventually become []string{"incubating", "experimental"}. The explicit
selection of the implementation for benchmarking in
EnableAllocators("experimental") then becomes EnableAllocators("incubating").
Documentation
¶
Index ¶
Constants ¶
This section is empty.
Variables ¶
Functions ¶
func GenerateShareID ¶
Types ¶
type AllocatedState ¶
type AllocatedState struct {
AllocatedDevices sets.Set[DeviceID]
AggregatedCapacity ConsumedCapacityCollection
}
AllocatedState packs information of allocated devices which is gathered from allocated resource claims.
type Allocator ¶
type Allocator interface {
Allocate(ctx context.Context, node *v1.Node, claims []*resourceapi.ResourceClaim) (finalResult []resourceapi.AllocationResult, finalErr error)
}
Allocator is intentionally not documented here. See the main package for docs.
This interface is also broader than the public one.
type AllocatorExtended ¶
type AllocatorExtended interface {
// Stats shows statistics from the allocation process.
// May return nil if not implemented.
GetStats() Stats
}
AllocatorExtended is an optional interface. Not all variants implement it.
type ConsumedCapacity ¶
type ConsumedCapacity map[resourceapi.QualifiedName]*resource.Quantity
ConsumedCapacity defines consumable capacity values
func NewConsumedCapacity ¶
func NewConsumedCapacity() ConsumedCapacity
NewConsumedCapacity initiates a new map of consumable capacity values
func (ConsumedCapacity) Add ¶
func (s ConsumedCapacity) Add(addedCapacity ConsumedCapacity)
Add adds quantity to corresponding consumable capacity, and creates a new entry if no capacity created yet.
func (ConsumedCapacity) Clone ¶
func (s ConsumedCapacity) Clone() ConsumedCapacity
Clone makes a copy of consumed capacity values
func (ConsumedCapacity) Empty ¶
func (s ConsumedCapacity) Empty() bool
Empty return true if all quantity is zero.
func (ConsumedCapacity) Sub ¶
func (s ConsumedCapacity) Sub(subtractedCapacity ConsumedCapacity)
Sub subtracts quantity, and ignore if no capacity entry found.
type ConsumedCapacityCollection ¶
type ConsumedCapacityCollection map[DeviceID]ConsumedCapacity
ConsumedCapacityCollection collects consumable capacity values of each device
func NewConsumedCapacityCollection ¶
func NewConsumedCapacityCollection() ConsumedCapacityCollection
NewConsumedCapacity initiates a new map of device's consumable capacity values
func (ConsumedCapacityCollection) Clone ¶
func (c ConsumedCapacityCollection) Clone() ConsumedCapacityCollection
Clone makes a copy of ConsumedCapacity of each capacity.
func (ConsumedCapacityCollection) Insert ¶
func (c ConsumedCapacityCollection) Insert(cap DeviceConsumedCapacity)
Insert adds a new allocated capacity to the collection.
func (ConsumedCapacityCollection) Remove ¶
func (c ConsumedCapacityCollection) Remove(cap DeviceConsumedCapacity)
Remove removes an allocated capacity from the collection.
type DeviceClassLister ¶
type DeviceClassLister interface {
// List returns a list of all DeviceClasses.
List() ([]*resourceapi.DeviceClass, error)
// Get returns the DeviceClass with the given className.
Get(className string) (*resourceapi.DeviceClass, error)
}
type DeviceConsumedCapacity ¶
type DeviceConsumedCapacity struct {
DeviceID
ConsumedCapacity
}
DeviceConsumedCapacity contains consumed capacity result within device allocation.
func NewDeviceConsumedCapacity ¶
func NewDeviceConsumedCapacity(deviceID DeviceID, consumedCapacity map[resourceapi.QualifiedName]resource.Quantity) DeviceConsumedCapacity
NewDeviceConsumedCapacity creates DeviceConsumedCapacity instance from device ID and its consumed capacity.
func (DeviceConsumedCapacity) Clone ¶
func (a DeviceConsumedCapacity) Clone() DeviceConsumedCapacity
Clone makes a copy of DeviceConsumedCapacity.
func (DeviceConsumedCapacity) String ¶
func (a DeviceConsumedCapacity) String() string
String returns formatted device ID.
type DeviceID ¶
type DeviceID struct {
Driver, Pool, Device draapi.UniqueString
}
func MakeDeviceID ¶
type Features ¶
type Features struct {
AdminAccess bool
ConsumableCapacity bool
DeviceBindingAndStatus bool
DeviceTaints bool
PartitionableDevices bool
PrioritizedList bool
}
Features control optional functionality during ResourceClaim allocation. Each entry must correspond to at least one control flow change. Entries can be removed when the control flow change is no longer necessary (= feature is considered stable and always enabled).
This often corresponds to feature gates, but not always: if a KEP implementation depends on a set of feature gates, then a single entry here should control whether that implementation is active.
func (Features) Set ¶
Set returns all features which are set to true. The names of the features happen to match the Kubernetes feature gates where applicable. Plain strings are used because not all allocator features necessarily have to be Kubernetes feature gates and this package must not depend on those definitions.
type SharedDeviceID ¶
type SharedDeviceID struct {
func MakeSharedDeviceID ¶
func MakeSharedDeviceID(deviceID DeviceID, shareID *types.UID) SharedDeviceID
MakeSharedDeviceID creates a SharedDeviceID by extending MakeDeviceID with shareID.
func (SharedDeviceID) String ¶
func (d SharedDeviceID) String() string
Source Files
Directories