[
https://issues.apache.org/jira/browse/SPARK-23082?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Oz Ben-Ami updated SPARK-23082:
-------------------------------
Description:
In YARN, we can use spark.yarn.am.nodeLabelExpression to submit the Spark
driver to a different set of nodes from its executors. In Kubernetes, we can
specify spark.kubernetes.node.selector.[labelKey], but we can't use separate
options for the driver and executors.
This would be useful for the particular use case where executors can go on more
ephemeral nodes (eg, with cluster autoscaling, or preemptible/spot instances),
but the driver should use a more persistent machine.
The required change would be minimal, essentially just using different config
keys in
[https://github.com/apache/spark/blob/master/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/k8s/submit/steps/BasicDriverConfigurationStep.scala#L90]
and
[https://github.com/apache/spark/blob/0b2eefb674151a0af64806728b38d9410da552ec/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/scheduler/cluster/k8s/ExecutorPodFactory.scala#L73]
instead of KUBERNETES_NODE_SELECTOR_PREFIX for both.
was:
In YARN, we can use spark.yarn.am.nodeLabelExpression to submit the Spark
driver to a different set of nodes from its executors. In Kubernetes, we can
specify spark.kubernetes.node.selector.[labelKey], but we can't use separate
options for the driver and executors. This would be useful for the particular
use case where executors can go on more ephemeral nodes (eg, with cluster
autoscaling, or preemptible/spot instances), but the driver should use a more
persistent machine.
The required change would be minimal, essentially just using different config
keys in
[https://github.com/apache/spark/blob/master/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/k8s/submit/steps/BasicDriverConfigurationStep.scala#L90]
and
[https://github.com/apache/spark/blob/0b2eefb674151a0af64806728b38d9410da552ec/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/scheduler/cluster/k8s/ExecutorPodFactory.scala#L73]
instead of KUBERNETES_NODE_SELECTOR_PREFIX for both.
> Allow separate node selectors for driver and executors in Kubernetes
> --------------------------------------------------------------------
>
> Key: SPARK-23082
> URL: https://issues.apache.org/jira/browse/SPARK-23082
> Project: Spark
> Issue Type: Improvement
> Components: Kubernetes, Spark Submit
> Affects Versions: 2.2.0, 2.3.0
> Reporter: Oz Ben-Ami
> Priority: Minor
>
> In YARN, we can use spark.yarn.am.nodeLabelExpression to submit the Spark
> driver to a different set of nodes from its executors. In Kubernetes, we can
> specify spark.kubernetes.node.selector.[labelKey], but we can't use separate
> options for the driver and executors.
> This would be useful for the particular use case where executors can go on
> more ephemeral nodes (eg, with cluster autoscaling, or preemptible/spot
> instances), but the driver should use a more persistent machine.
> The required change would be minimal, essentially just using different config
> keys in
> [https://github.com/apache/spark/blob/master/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/k8s/submit/steps/BasicDriverConfigurationStep.scala#L90]
> and
> [https://github.com/apache/spark/blob/0b2eefb674151a0af64806728b38d9410da552ec/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/scheduler/cluster/k8s/ExecutorPodFactory.scala#L73]
> instead of KUBERNETES_NODE_SELECTOR_PREFIX for both.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
