[
https://issues.apache.org/jira/browse/SPARK-24288?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16477747#comment-16477747
]
Maryann Xue commented on SPARK-24288:
-------------------------------------
Thank you, [~cloud_fan] and [~TomaszGaweda] for the input! Here's some ideas
from my side:
Basically the question comes down to whether 1) avoid predicate push-down to
data source level or 2) avoid predicate push-down (or other push-down
optimization) anywhere in the logical operator tree.
For 1) we could have the first two solutions [~cloud_fan] has proposed or the
hint solution [~smilegator] has proposed.
For 2) we would use a logical barrier as I have mentioned earlier.
I'm in favor of the hint approach coz it would be an easy interface to use,
less extendable maybe. It can be used for the entire query or with DataFrame
interface as well. For example, like:
{{spark.sql("select /*+ NO_PRED_PUSHDOWN */ from foobar where name like 'x%'")}}
or
{{spark.jdbc(...).where("name like 'x%'").hint("NO_PRED_PUSHDOWN")}}
On the other hand, 2) is more powerful, for example, can be used to push part
of the predicates down or to push certain while not other operators down. And
like I said, it might not require a whole lot of changes. I am attaching an
experiment patch here to as a quick illustration.
> Enable preventing predicate pushdown
> ------------------------------------
>
> Key: SPARK-24288
> URL: https://issues.apache.org/jira/browse/SPARK-24288
> Project: Spark
> Issue Type: New Feature
> Components: SQL
> Affects Versions: 2.3.0
> Reporter: Tomasz Gawęda
> Priority: Major
>
> Issue discussed on Mailing List:
> [http://apache-spark-developers-list.1001551.n3.nabble.com/Preventing-predicate-pushdown-td23976.html]
> While working with JDBC datasource I saw that many "or" clauses with
> non-equality operators causes huge performance degradation of SQL query
> to database (DB2). For example:
> val df = spark.read.format("jdbc").(other options to parallelize
> load).load()
> df.where(s"(date1 > $param1 and (date1 < $param2 or date1 is null) or x
> > 100)").show() // in real application whose predicates were pushed
> many many lines below, many ANDs and ORs
> If I use cache() before where, there is no predicate pushdown of this
> "where" clause. However, in production system caching many sources is a
> waste of memory (especially is pipeline is long and I must do cache many
> times).There are also few more workarounds, but it would be great if Spark
> will support preventing predicate pushdown by user.
>
> For example: df.withAnalysisBarrier().where(...) ?
>
> Note, that this should not be a global configuration option. If I read 2
> DataFrames, df1 and df2, I would like to specify that df1 should not have
> some predicates pushed down, but some may be, but df2 should have all
> predicates pushed down, even if target query joins df1 and df2. As far as I
> understand Spark optimizer, if we use functions like `withAnalysisBarrier`
> and put AnalysisBarrier explicitly in logical plan, then predicates won't be
> pushed down on this particular DataFrames and PP will be still possible on
> the second one.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
