Github user jsoltren commented on a diff in the pull request:
https://github.com/apache/spark/pull/16650#discussion_r98078786
--- Diff:
core/src/main/scala/org/apache/spark/scheduler/BlacklistTracker.scala ---
@@ -173,6 +174,24 @@ private[scheduler] class BlacklistTracker (
listenerBus.post(SparkListenerExecutorBlacklisted(now, exec,
newTotal))
executorIdToFailureList.remove(exec)
updateNextExpiryTime()
+ // Add executor to blacklist before attempting to kill it. This
allows a scheduler backend
+ // to immediately fail to allocate resources on this executor,
since killing could be
+ // asynchronous.
--- End diff --
I put this comment here to avoid a race between an individual executor
being blacklisted, and being scheduled for killing. The comment makes more
sense in the context of an individual executor, and not an entire node.
What you're talking about is something different, and more complex: the
need to - explicitly - blacklist all the executors on a node before marking
that node for killing.
For example, consider hostA with executors 1, 2, and 3. We can blacklist 1,
and then 2, and then hostA, but 3 is never explicitly blacklisted. The
`nodeBlacklist` case needs to blacklist 3, which means we need some explicit
mapping between hosts and the executors on that host in BlacklistTracker. To
date we've been trying to avoid exactly this to "keep the BlacklistTracker
simple".
---
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]
