[
https://issues.apache.org/jira/browse/MAHOUT-1597?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14071023#comment-14071023
]
ASF GitHub Bot commented on MAHOUT-1597:
----------------------------------------
Github user dlyubimov commented on the pull request:
https://github.com/apache/mahout/pull/33#issuecomment-49809300
the point is there has to be diagnostics propagating thru DAG to help
rewrites based on initial quick summaries of the checkpoints.
E.g. in initial version there was also a non-zero element estimate that (I
hoped) would help optimizer to take certain rewrite decisions. I removed it as
it was too much for the first step. But if cost optimizations ever become a
larger part of this, operators need to be able to carry optimizer specific
information .
RDDInput stuff on the other hand is merely an "either/or" rdd (either
blockifed or row-wise). As such it is the sole final product of the optimizer.
Result can't carry helper information pertinent to building result tree, I feel
that would be a bad optimizer architecture.
> A + 1.0 (element-wise scala operation) gives wrong result if rdd is missing
> rows, Spark side
> --------------------------------------------------------------------------------------------
>
> Key: MAHOUT-1597
> URL: https://issues.apache.org/jira/browse/MAHOUT-1597
> Project: Mahout
> Issue Type: Bug
> Affects Versions: 0.9
> Reporter: Dmitriy Lyubimov
> Assignee: Dmitriy Lyubimov
> Fix For: 1.0
>
>
> {code}
> // Concoct an rdd with missing rows
> val aRdd: DrmRdd[Int] = sc.parallelize(
> 0 -> dvec(1, 2, 3) ::
> 3 -> dvec(3, 4, 5) :: Nil
> ).map { case (key, vec) => key -> (vec: Vector)}
> val drmA = drmWrap(rdd = aRdd)
> val controlB = inCoreA + 1.0
> val drmB = drmA + 1.0
> (drmB -: controlB).norm should be < 1e-10
> {code}
> should not fail.
> it was failing due to elementwise scalar operator only evaluates rows
> actually present in dataset.
> In case of Int-keyed row matrices, there are implied rows that yet may not be
> present in RDD.
> Our goal is to detect the condition and evaluate missing rows prior to
> physical operators that don't work with missing implied rows.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
