[
https://issues.apache.org/jira/browse/SPARK-12030?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15033321#comment-15033321
]
Nong Li commented on SPARK-12030:
---------------------------------
I think I tracked it down. The bug is from this PR which exposed another bug.
SPARK-7542[SQL] Support off-heap index/sort buffer
https://github.com/apache/spark/pull/9477
The issue I think is in
https://github.com/apache/spark/blob/master/unsafe/src/main/java/org/apache/spark/unsafe/Platform.java#L108
which does not support overlapping regions. For example,
If the src region is [100 - 1000) and the dest region is [500-1500) and the
copy size is 100, the first copy of 100 bytes would overwrite the bytes at
offset 500. This needs a check to see that the src region is before the dest
and run this loop backwards I think.
This only hits on larger data sets when a single copy exceeds the threshold
(1M).
> Incorrect results when aggregate joined data
> --------------------------------------------
>
> Key: SPARK-12030
> URL: https://issues.apache.org/jira/browse/SPARK-12030
> Project: Spark
> Issue Type: Bug
> Components: SQL
> Affects Versions: 1.6.0
> Reporter: Maciej Bryński
> Priority: Blocker
> Attachments: spark.jpg, t1.tar.gz, t2.tar.gz
>
>
> I have following issue.
> I created 2 dataframes from JDBC (MySQL) and joined them (t1 has fk1 to t2)
> {code}
> t1 = sqlCtx.read.jdbc("jdbc:mysql://XXX", t1, id1, 0, size1, 200).cache()
> t2 = sqlCtx.read.jdbc("jdbc:mysql://XXX", t2).cache()
> joined = t1.join(t2, t1.fk1 == t2.id2, "left_outer")
> {code}
> Important: both table are cached, so results should be the same on every
> query.
> Then I did come counts:
> {code}
> t1.count() -> 5900729
> t1.registerTempTable("t1")
> sqlCtx.sql("select distinct(id1) from t1").count() -> 5900729
> t2.count() -> 54298
> joined.count() -> 5900729
> {code}
> And here magic begins - I counted distinct id1 from joined table
> {code}
> joined.registerTempTable("joined")
> sqlCtx.sql("select distinct(id1) from joined").count()
> {code}
> Results varies *(are different on every run)* between 5899000 and
> 5900000 but never are equal to 5900729.
> In addition. I did more queries:
> {code}
> sqlCtx.sql("select id1, count(*) from joined group by id1 having count(*) >
> 1").collect()
> {code}
> This gives some results but this query return *1*
> {code}
> len(sqlCtx.sql("select * from joined where id1 = result").collect())
> {code}
> What's wrong ?
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
