Skip to content

Latest commit

 

History

History
112 lines (98 loc) · 5.67 KB

affinityassistants.md

File metadata and controls

112 lines (98 loc) · 5.67 KB

Affinity Assistants

Affinity Assistant is a feature to coschedule PipelineRun pods to the same node based on kubernetes pod affinity so that it possible for the taskruns to execute parallel while sharing volume. Available Affinity Assistant Modes are coschedule workspaces, coschedule pipelineruns, isolate pipelinerun and disabled.

🌱 coschedule pipelineruns and isolate pipelinerun modes are alpha features. coschedule workspaces is a stable feature

  • coschedule workspaces - When a PersistentVolumeClaim is used as volume source for a Workspace in a PipelineRun, all TaskRun pods within the PipelineRun that share the Workspace will be scheduled to the same Node. (Note: Only one pvc-backed workspace can be mounted to each TaskRun in this mode.)

  • coschedule pipelineruns - All TaskRun pods within the PipelineRun will be scheduled to the same Node.

  • isolate pipelinerun - All TaskRun pods within the PipelineRun will be scheduled to the same Node, and only one PipelineRun is allowed to run on a node at a time.

  • disabled - The Affinity Assistant is disabled. No pod coscheduling behavior.

This means that Affinity Assistant is incompatible with other affinity rules configured for the TaskRun pods (i.e. other affinity rules specified in custom PodTemplate will be overwritten by Affinity Assistant). If the PipelineRun has a custom PodTemplate configured, the NodeSelector and Tolerations fields will also be set on the Affinity Assistant pod. The Affinity Assistant is deleted when the PipelineRun is completed.

Currently, the Affinity Assistant Modes can be configured by the disable-affinity-assistant and coschedule feature flags. The disable-affinity-assistant feature flag is now deprecated and will be removed in release v0.60. At the time, the Affinity Assistant Modes will be only determined by the coschedule feature flag.

The following chart summarizes the Affinity Assistant Modes with different combinations of the disable-affinity-assistant and coschedule feature flags during migration (when both feature flags are present) and after the migration (when only the coschedule flag is present).

disable-affinity-assistant coschedule behavior during migration behavior after migration
false (default) disabled N/A: invalid disabled
false (default) workspaces (default) coschedule workspaces coschedule workspaces
false (default) pipelineruns N/A: invalid coschedule pipelineruns
false (default) isolate-pipelinerun N/A: invalid isolate pipelinerun
true disabled disabled disabled
true workspaces (default) disabled coschedule workspaces
true pipelineruns coschedule pipelineruns coschedule pipelineruns
true isolate-pipelinerun isolate pipelinerun isolate pipelinerun

Note: For users who previously accepted the default behavior (disable-affinity-assistant: false) but now want one of the new features, you need to set disable-affinity-assistant to "true" and then turn on the new behavior by setting the coschedule flag. For users who previously disabled the affinity assistant but want one of the new features, just set the coschedule flag accordingly.

Note: Affinity Assistant use Inter-pod affinity and anti-affinity that require substantial amount of processing which can slow down scheduling in large clusters significantly. We do not recommend using the affinity assistant in clusters larger than several hundred nodes

Note: Pod anti-affinity requires nodes to be consistently labelled, in other words every node in the cluster must have an appropriate label matching topologyKey. If some or all nodes are missing the specified topologyKey label, it can lead to unintended behavior.

Note: Any time during the execution of a pipelineRun, if the node with a placeholder Affinity Assistant pod and the taskRun pods sharing a workspace is cordoned or disabled for scheduling anything new (tainted), the pipelineRun controller deletes the placeholder pod. The taskRun pods on a cordoned node continues running until completion. The deletion of a placeholder pod triggers creating a new placeholder pod on any available node such that the rest of the pipelineRun can continue without any disruption until it finishes.