Skip to content

Latest commit

 

History

History
111 lines (75 loc) · 7.74 KB

scheduler.md

File metadata and controls

111 lines (75 loc) · 7.74 KB
Raw

Gardener Scheduler

The Gardener Scheduler is in essence a controller that watches newly created shoots and assigns a seed cluster to them. Conceptually, the task of the Gardener Scheduler is very similar to the task of the Kubernetes Scheduler: finding a seed for a shoot instead of a node for a pod.

Either the scheduling strategy or the shoot cluster purpose hereby determines how the scheduler is operating. The following sections explain the configuration and flow in greater detail.

Why is the Gardener Scheduler needed?

1. Decoupling

Previously, an admission plugin in the Gardener API server conducted the scheduling decisions. This implies changes to the API server whenever adjustments of the scheduling are needed. Decoupling the API server and the scheduler comes with greater flexibility to develop these components independently from each other.

2. Extensibility

It should be possible to easily extend and tweak the scheduler in the future. Possibly, similar to the Kubernetes scheduler, hooks could be provided which influence the scheduling decisions. It should be also possible to completely replace the standard Gardener Scheduler with a custom implementation.

Algorithm overview

The following sequence describes the steps involved to determine a seed candidate:

  1. Determine usable seeds with "usable" defined as follows:
    • no .metadata.deletionTimestamp
    • .spec.settings.scheduling.visible is true
    • conditions Bootstrapped, GardenletReady, BackupBucketsReady (if available) are true
  2. Filter seeds:
    • matching .spec.seedSelector in CloudProfile used by the Shoot
    • matching .spec.seedSelector in Shoot
    • having no network intersection with the Shoot's networks (due to the VPN connectivity between seeds and shoots their networks must be disjoint)
    • having .spec.settings.shootDNS.enabled=false (only if the shoot specifies a DNS domain or does not use the unmanaged DNS provider)
    • whose taints (.spec.taints) are tolerated by the Shoot (.spec.tolerations)
    • whose capacity for shoots would not be exceeded if the shoot is scheduled onto the seed, see Ensuring seeds capacity for shoots is not exceeded
  3. Apply active strategy e.g., Minimal Distance strategy
  4. Choose least utilized seed, i.e., the one with the least number of shoot control planes, will be the winner and written to the .spec.seedName field of the Shoot.

Configuration

The Gardener Scheduler configuration has to be supplied on startup. It is a mandatory and also the only available flag. Here is an example scheduler configuration.

Most of the configuration options are the same as in the Gardener Controller Manager (leader election, client connection, ...). However, the Gardener Scheduler on the other hand does not need a TLS configuration, because there are currently no webhooks configurable.

Strategies

The scheduling strategy is defined in the candidateDeterminationStrategy of the scheduler's configuration and can have the possible values SameRegion and MinimalDistance. The SameRegion strategy is the default strategy.

  1. Same Region strategy

    The Gardener Scheduler reads the spec.provider.type and .spec.region fields from the Shoot resource. It tries to find a seed that has the identical .spec.provider.type and .spec.provider.region fields set. If it cannot find a suitable seed, it adds an event to the shoot stating, that it is unschedulable.

  2. Minimal Distance strategy

    The Gardener Scheduler tries to find a valid seed with minimal distance to the shoot's intended region. The distance is calculated based on the Levenshtein distance of the region. Therefore the region name is split into a base name and an orientation. Possible orientations are north, south, east, west and central. The distance then is twice the Levenshtein distance of the region's base name plus a correction value based on the orientation and the provider.

    If the orientations of shoot and seed candidate match, the correction value is 0, if they differ it is 2 and if either the seed's or the shoot's region does not have an orientation it is 1. If the provider differs the correction value is additionally incremented by 2.

    Because of this a matching region with a matching provider is always prefered.

In order to put the scheduling decision into effect, the scheduler sends an update request for the Shoot resource to the API server. After validation, the Gardener Aggregated API server updates the shoot to have the spec.seedName field set. Subsequently, the Gardenlet picks up and starts to create the cluster on the specified seed.

  1. Special handling based on shoot cluster purpose

Every shoot cluster can have a purpose that describes what the cluster is used for, and also influences how the cluster is setup (see this document for more information).

In case the shoot has the testing purpose then the scheduler only reads the .spec.provider.type from the Shoot resource and tries to find a Seed that has the identical .spec.provider.type. The region does not matter, i.e., testing shoots may also be scheduled on a seed in a complete different region if it is better for balancing the whole Gardener system.

seedSelector field in the Shoot specification

Similar to the .spec.nodeSelector field in Pods, the Shoot specification has an optional .spec.seedSelector field. It allows the user to provide a label selector that must match the labels of Seeds in order to be scheduled to one of them. The labels on Seeds are usually controlled by Gardener administrators/operators - end users cannot add arbitrary labels themselves. If provided, the Gardener Scheduler will only consider those seeds as "suitable" whose labels match those provided in the .spec.seedSelector of the Shoot.

By default only seeds with the same provider than the shoot are selected. By adding a providerTypes field to the seedSelector a dedicated set of possible providers (* means all provider types) can be selected.

Ensuring seeds capacity for shoots is not exceeded

Seeds have a practical limit of how many shoots they can accommodate. Exceeding this limit is undesirable as the system performance will be noticeably impacted. Therefore, the scheduler ensures that a seed's capacity for shoots is not exceeded by taking into account a maximum number of shoots that can be scheduled onto a seed.

This mechanism works as follows:

  • The gardenlet is configured with certain resources and their total capacity (and, for certain resources, the amount reserved for Gardener), see /example/20-componentconfig-gardenlet.yaml. Currently, the only such resource is the maximum number of shoots that can be scheduled onto a seed.
  • The gardenlet seed controller updates the capacity and allocatable fields in Seed status with the capacity of each resource and how much of it is actually available to be consumed by shoots. The allocatable value of a resource is equal to capacity minus reserved.
  • When scheduling shoots, the scheduler filters out all candidate seeds whose allocatable capacity for shoots would be exceeded if the shoot is scheduled onto the seed.

Failure to determine a suitable seed

In case the scheduler fails to find a suitable seed, the operation is being retried with exponential backoff.

Current Limitation / Future Plans

  • Azure has unfortunately a geographically non-hierarchical naming pattern and does not start with the continent. This is the reason why we will exchange the implementation of the MinimalDistance strategy with a more suitable one in the future.