[
https://issues.apache.org/jira/browse/SPARK-24904?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16555578#comment-16555578
]
Shay Elbaz commented on SPARK-24904:
------------------------------------
[~mgaido] Technically you *can* that, you just need an additional shuffle
(after the map side join) to fill in the missing rows as you mentioned. And
since the current implementation already shuffles, I don't see how it makes
sense to involve the entire big table in the shuffle. Instead, Spark could do
the following:
# Broadcast the small table.
# Just link inner join, load the big table and hash-join.
The output is (expected to be) very small compared to the big table.
# Keep the small table broadcasted, and shuffle the results from last stage
(say, sort-merge).
# Now on each task, fill in missing rows from the broadcasted table. This is
trivial if using sort-merge and the broadcasted table is just another partition
to merge.
As I mentioned in the description, this is can be achieved by the user using 2
joins, but shouldn't Spark offer this by default? Needless to say how sub
optimal the current implementation is compared to the above plan. Am I missing
something?
> Join with broadcasted dataframe causes shuffle of redundant data
> ----------------------------------------------------------------
>
> Key: SPARK-24904
> URL: https://issues.apache.org/jira/browse/SPARK-24904
> Project: Spark
> Issue Type: Question
> Components: SQL
> Affects Versions: 2.1.2
> Reporter: Shay Elbaz
> Priority: Minor
>
> When joining a "large" dataframe with broadcasted small one, and join-type is
> on the small DF side (see right-join below), the physical plan does not
> include broadcasting the small table. But when the join is on the large DF
> side, the broadcast does take place. Is there a good reason for this? In the
> below example it sure doesn't make any sense to shuffle the entire large
> table:
>
> {code:java}
> val small = spark.range(1, 10)
> val big = spark.range(1, 1 << 30)
> .withColumnRenamed("id", "id2")
> big.join(broadcast(small), $"id" === $"id2", "right")
> .explain
> //OUTPUT:
> == Physical Plan ==
> SortMergeJoin [id2#16307L], [id#16310L], RightOuter
> :- *Sort [id2#16307L ASC NULLS FIRST], false, 0
> : +- Exchange hashpartitioning(id2#16307L, 1000)
> : +- *Project [id#16304L AS id2#16307L]
> : +- *Range (1, 1073741824, step=1, splits=Some(600))
> +- *Sort [id#16310L ASC NULLS FIRST], false, 0
> +- Exchange hashpartitioning(id#16310L, 1000)
> +- *Range (1, 10, step=1, splits=Some(600))
> {code}
> As a workaround, users need to perform inner instead of right join, and then
> join the result back with the small DF to fill the missing rows.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
