[
https://issues.apache.org/jira/browse/SPARK-34694?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17301701#comment-17301701
]
Chen Zou commented on SPARK-34694:
----------------------------------
Hi Hyukjin,
I think the design you described would work.
But the current org.apache.spark.sql.sources.Filter isn't built under the
assumption that the 'value' parameter could be a column reference.
e.g. the findReferences member function does not consider value being a column
references.
{quote} protected def findReferences(value: Any): Array[String] = value match {
case f: Filter => f.references
case _ => Array.empty
}{quote}
And this is probably why
org.apache.spark.sql.execution.datasources.v2.PushDownUtils would not push the
cross-column filters down to data sources.
The end result is that cross-column filters don't get pushed down, from stderr
of a spark job doing TPC-H Q12:
21/03/10 16:56:16.266 INFO V2ScanRelationPushDown:
Pushing operators to
lineitem@file:///home/colouser51/udpstorage/tpch/tbl_s1e1/lineitem
Pushed Filters: Or(EqualTo(l_shipmode,MAIL),EqualTo(l_shipmode,SHIP)),
GreaterThanOrEqual(l_receiptdate,1994-01-01), LessThan(l_receiptdate,1995-01-01)
Post-Scan Filters: (l_commitdate#11 < l_receiptdate#12),(l_shipdate#10 <
l_commitdate#11)
Output: l_orderkey#0, l_shipdate#10, l_commitdate#11, l_receiptdate#12,
l_shipmode#14
Regards,
Chen
> Improve Spark SQL Source Filter to allow pushdown of filters span multiple
> columns
> ----------------------------------------------------------------------------------
>
> Key: SPARK-34694
> URL: https://issues.apache.org/jira/browse/SPARK-34694
> Project: Spark
> Issue Type: Improvement
> Components: Spark Core
> Affects Versions: 3.0.0, 3.0.1, 3.0.2, 3.1.0, 3.1.1
> Reporter: Chen Zou
> Priority: Minor
>
> The current org.apache.spark.sql.sources.Filter abstract class only allows
> pushdown of filters on single column or sum of products of multiple such
> single-column filters.
> Filters on multiple columns cannot be pushed down through this Filter
> subclass to source, e.g. from TPC-H benchmark on lineitem table:
> (l_commitdate#11 < l_receiptdate#12)
> (l_shipdate#10 < l_commitdate#11)
>
> The current design probably originates from the point that columnar source
> has a hard time supporting these cross-column filters. But with batching
> implemented in columnar sources, they can still support cross-column filters.
> This issue tries to open up discussion on a more general Filter interface to
> allow pushing down cross-column filters.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
