[
https://issues.apache.org/jira/browse/SPARK-18735?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Sean Owen resolved SPARK-18735.
-------------------------------
Resolution: Invalid
(Please ask questions on the mailing list)
Two reasons:
1) if the computation that uses the broadcast isn't triggered within the loop
iteration, then you will be removing the broadcast before it's used
2) destroy()-ing even after the broadcast has definitely been used to compute,
say, a cached RDD, is a problem because the cached RDD might still be
recomputed and a destroyed broadcast can't be used. unpersist() is more correct
in this case.
> Why don't we destroy the broadcast variable after each iteration?
> -----------------------------------------------------------------
>
> Key: SPARK-18735
> URL: https://issues.apache.org/jira/browse/SPARK-18735
> Project: Spark
> Issue Type: Improvement
> Components: MLlib
> Affects Versions: 2.0.2
> Reporter: Jianfei Wang
>
> I think we should destroy the broadcast variable bcWeights explicitly in
> spark.mllib.GradientDescent.runMiniBatchSGD
> see the code below:
> while (!converged && i <= numIterations) {
> val start = System.nanoTime()
> val bcWeights = data.context.broadcast(weights)
> // some other code
> }
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
