[
https://issues.apache.org/jira/browse/SPARK-32708?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17185523#comment-17185523
]
Mingjia Liu edited comment on SPARK-32708 at 8/27/20, 12:01 AM:
----------------------------------------------------------------
This difference happened at ReuseExchange rule apply.
The direct reason is : equals function of DataSourceV2ScanExec returns 'false'
as comparing [6] [5] VS [2][1] in DataSourceV2 plan.
But the actual cause is : DataSourceV2ScanExec.pushedFilters are defined as
array of Expressions whose equal function has expression_id in its scope.
So for example, *isnotnull(d_day_name#22364)* is not considered equal to
*isnotnull(d_day_name#22420)*
was (Author: mingjial):
This difference happened at ReuseExchange rule apply.
The direct reason is : equals function of DataSourceV2ScanExec returns 'false'
as comparing [6] [5] VS [2][1] in DataSourceV2 plan.
But the actual cause is : DataSourceV2ScanExec.pushedFilters are defined as
array of Expressions whose equal function has expression_id in its scope.
So for example,
isnotnull(d_day_name#22364)
is not considered equal to
isnotnull(d_day_name#22420)
> Query optimization fails to reuse exchange with DataSourceV2
> ------------------------------------------------------------
>
> Key: SPARK-32708
> URL: https://issues.apache.org/jira/browse/SPARK-32708
> Project: Spark
> Issue Type: Bug
> Components: SQL
> Affects Versions: 2.4.0, 2.4.1, 2.4.2, 2.4.3, 2.4.4, 2.4.5, 2.4.6, 2.4.7
> Reporter: Mingjia Liu
> Priority: Major
>
> Repro query:
> {code:java}
> spark.conf.set("spark.sql.exchange.reuse","true")
> spark.read.format('parquet').load('gs://dataproc-kokoro-tests-us-central1/tpcds/1G/parquet/date_dim')
> #spark.read.format("com.google.cloud.spark.bigquery.v2.BigQueryDataSourceV2").option("table",'tpcds_1G.date_dim').load()
>
> df.createOrReplaceTempView(table)
>
> df = spark.sql("""
> WITH t1 AS (
> SELECT
> d_year, d_month_seq
> FROM (
> SELECT t1.d_year , t2.d_month_seq
> FROM
> date_dim t1
> cross join
> date_dim t2
> where t1.d_day_name = "Monday" and t1.d_fy_year > 2000
> and t2.d_day_name = "Monday" and t2.d_fy_year > 2000
> )
> GROUP BY d_year, d_month_seq)
>
> SELECT
> prev_yr.d_year AS prev_year, curr_yr.d_year AS year, curr_yr.d_month_seq
> FROM t1 curr_yr cross join t1 prev_yr
> WHERE curr_yr.d_year=2002 AND prev_yr.d_year=2002
> ORDER BY d_month_seq
> LIMIT 100
>
> """)
> df.explain()
> #df.show()
> {code}
>
> *The above query has different plans with Parquet and DataSourceV2. Both
> plans are correct tho. However, the DataSourceV2 plan is less optimized :*
> *Sub-plan [5-7] is exactly the same as sub-plan [1-3]( Aggregate on BHJed
> dataset of two tables that are filtered, projected the same way).*
> *Therefore, in the below parquet plan, exchange that happens after [1-3] is
> reused to replace [5-6].*
> *However, the DataSourceV2 plan failed to do so.*
>
> Parquet:
> {code:java}
> == Physical Plan ==
> TakeOrderedAndProject(limit=100, orderBy=[d_month_seq#21456L ASC NULLS
> FIRST], output=[prev_year#21451L,year#21452L,d_month_seq#21456L])
> +- *(9) Project [d_year#21487L AS prev_year#21451L, d_year#20481L AS
> year#21452L, d_month_seq#21456L]
> +- CartesianProduct
> :- *(4) HashAggregate(keys=[d_year#20481L, d_month_seq#21456L],
> functions=[])
> : +- Exchange hashpartitioning(d_year#20481L, d_month_seq#21456L, 200)
> : +- *(3) HashAggregate(keys=[d_year#20481L, d_month_seq#21456L],
> functions=[])
> : +- BroadcastNestedLoopJoin BuildRight, Cross
> : :- *(1) Project [d_year#20481L]
> : : +- *(1) Filter (((((isnotnull(d_year#20481L) &&
> isnotnull(d_day_name#20489)) && isnotnull(d_fy_year#20486L)) &&
> (d_day_name#20489 = Monday)) && (d_fy_year#20486L > 2000)) && (d_year#20481L
> = 2002))
> : : +- *(1) FileScan parquet
> [d_year#20481L,d_fy_year#20486L,d_day_name#20489] Batched: true, Format:
> Parquet, Location:
> InMemoryFileIndex[gs://dataproc-kokoro-tests-us-central1/tpcds/1G/parquet/date_dim],
> PartitionFilters: [], PushedFilters: [IsNotNull(d_year),
> IsNotNull(d_day_name), IsNotNull(d_fy_year), EqualTo(d_day_name,Monday),
> Grea..., ReadSchema: struct<d_year:bigint,d_fy_year:bigint,d_day_name:string>
> : +- BroadcastExchange IdentityBroadcastMode
> : +- *(2) Project [d_month_seq#21456L]
> : +- *(2) Filter (((isnotnull(d_day_name#21467) &&
> isnotnull(d_fy_year#21464L)) && (d_day_name#21467 = Monday)) &&
> (d_fy_year#21464L > 2000))
> : +- *(2) FileScan parquet
> [d_month_seq#21456L,d_fy_year#21464L,d_day_name#21467] Batched: true, Format:
> Parquet, Location:
> InMemoryFileIndex[gs://dataproc-kokoro-tests-us-central1/tpcds/1G/parquet/date_dim],
> PartitionFilters: [], PushedFilters: [IsNotNull(d_day_name),
> IsNotNull(d_fy_year), EqualTo(d_day_name,Monday), GreaterThan(d_fy_year,2...,
> ReadSchema: struct<d_month_seq:bigint,d_fy_year:bigint,d_day_name:string>
> +- *(8) HashAggregate(keys=[d_year#21487L, d_month_seq#21540L],
> functions=[])
> +- ReusedExchange [d_year#21487L, d_month_seq#21540L], Exchange
> hashpartitioning(d_year#20481L, d_month_seq#21456L, 200){code}
>
> DataSourceV2:
> {code:java}
> == Physical Plan ==
> TakeOrderedAndProject(limit=100, orderBy=d_month_seq#22325L ASC NULLS FIRST,
> output=prev_year#22320L,year#22321L,d_month_seq#22325L)
> +- *(9) Project d_year#22356L AS prev_year#22320L, d_year#21696L AS
> year#22321L, d_month_seq#22325L
> +- CartesianProduct
> :- *(4) HashAggregate(keys=d_year#21696L, d_month_seq#22325L, functions=[])
> : +- Exchange hashpartitioning(d_year#21696L, d_month_seq#22325L, 200)
> : +- *(3) HashAggregate(keys=d_year#21696L, d_month_seq#22325L, functions=[])
> : +- BroadcastNestedLoopJoin BuildRight, Cross
> : :- *(1) Project d_year#21696L
> : : +- *(1) ScanV2 BigQueryDataSourceV2d_year#21696L (Filters:
> [isnotnull(d_day_name#21704), (d_day_name#21704 = Monday),
> isnotnull(d_fy_year#21701L), (d_fy_yea..., Options:
> [table=tpcds_1G.date_dim,paths=[]])
> : +- BroadcastExchange IdentityBroadcastMode
> : +- *(2) Project d_month_seq#22325L
> : +- *(2) ScanV2 BigQueryDataSourceV2d_month_seq#22325L (Filters:
> [isnotnull(d_day_name#22336), (d_day_name#22336 = Monday),
> isnotnull(d_fy_year#22333L), (d_fy_yea..., Options:
> [table=tpcds_1G.date_dim,paths=[]])
> +- *(8) HashAggregate(keys=d_year#22356L, d_month_seq#22409L, functions=[])
> +- Exchange hashpartitioning(d_year#22356L, d_month_seq#22409L, 200)
> +- *(7) HashAggregate(keys=d_year#22356L, d_month_seq#22409L, functions=[])
> +- BroadcastNestedLoopJoin BuildRight, Cross
> :- *(5) Project d_year#22356L
> : +- *(5) ScanV2 BigQueryDataSourceV2d_year#22356L (Filters:
> [isnotnull(d_day_name#22364), (d_day_name#22364 = Monday),
> isnotnull(d_fy_year#22361L), (d_fy_yea..., Options:
> [table=tpcds_1G.date_dim,paths=[]])
> +- BroadcastExchange IdentityBroadcastMode
> +- *(6) Project d_month_seq#22409L
> +- *(6) ScanV2 BigQueryDataSourceV2d_month_seq#22409L (Filters:
> [isnotnull(d_day_name#22420), (d_day_name#22420 = Monday),
> isnotnull(d_fy_year#22417L), (d_fy_yea..., Options:
> [table=tpcds_1G.date_dim,paths=[]]){code}
>
--
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
