Github user sryza commented on a diff in the pull request:
https://github.com/apache/spark/pull/6394#discussion_r31973076
--- Diff:
yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnAllocator.scala ---
@@ -225,12 +243,74 @@ private[yarn] class YarnAllocator(
logInfo(s"Will request $missing executor containers, each with
${resource.getVirtualCores} " +
s"cores and ${resource.getMemory} MB memory including
$memoryOverhead MB overhead")
- for (i <- 0 until missing) {
- val request = createContainerRequest(resource)
- amClient.addContainerRequest(request)
- val nodes = request.getNodes
- val hostStr = if (nodes == null || nodes.isEmpty) "Any" else
nodes.last
- logInfo(s"Container request (host: $hostStr, capability:
$resource)")
+ // Calculated the number of executors we expected to satisfy all the
preferred locality tasks
--- End diff --
One possible issue with the approach here is that our locality preferences
might shift when a stage completes and a new stage comes in. If we have a set
of outstanding requests, it would be nice if we could cancel requests for the
locations that we no longer need and place our requests on nodes that we now
care about.
One way we could achieve this is, each time we call `updateResourceRequsts`
remove all the unsatisfied container requests from amClient and resubmit
requests based on our current needs. I.e., after removal, missing would always
be `targetNumExecutors - numExecutorsRunning`.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
