[
https://issues.apache.org/jira/browse/SPARK-35985?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Wenchen Fan resolved SPARK-35985.
---------------------------------
Fix Version/s: 3.2.0
Resolution: Fixed
Issue resolved by pull request 33191
[https://github.com/apache/spark/pull/33191]
> File source V2 ignores partition filters when empty readDataSchema
> ------------------------------------------------------------------
>
> Key: SPARK-35985
> URL: https://issues.apache.org/jira/browse/SPARK-35985
> Project: Spark
> Issue Type: Bug
> Components: SQL
> Affects Versions: 3.0.0
> Reporter: Steven Aerts
> Assignee: Steven Aerts
> Priority: Major
> Fix For: 3.2.0
>
>
> A V2 datasource fails to rely on partition filters when it only wants to know
> how many entries there are, and is not interested of their context.
> So when the {{readDataSchema}} of the {{FileScan}} is empty, partition
> filters are not pushed down and all data is scanned.
> Some examples where this happens:
> {code:java}
> scala> spark.sql("SELECT count(*) FROM parq WHERE day=20210702").explain
> == Physical Plan ==
> *(2) HashAggregate(keys=[], functions=[count(1)])
> +- Exchange SinglePartition, ENSURE_REQUIREMENTS, [id=#136]
> +- *(1) HashAggregate(keys=[], functions=[partial_count(1)])
> +- *(1) Project
> +- *(1) Filter (isnotnull(day#68) AND (day#68 = 20210702))
> +- *(1) ColumnarToRow
> +- BatchScan[day#68] ParquetScan DataFilters: [], Format: parquet, Location:
> InMemoryFileIndex[file:/..., PartitionFilters: [], PushedFilers:
> [IsNotNull(day), EqualTo(day,20210702)], ReadSchema: struct<>, PushedFilters:
> [IsNotNull(day), EqualTo(day,20210702)]
> scala> spark.sql("SELECT input_file_name() FROM parq WHERE
> day=20210702").explain
> == Physical Plan ==
> *(1) Project [input_file_name() AS input_file_name()#131]
> +- *(1) Filter (isnotnull(day#68) AND (day#68 = 20210702))
> +- *(1) ColumnarToRow
> +- BatchScan[day#68] ParquetScan DataFilters: [], Format: parquet, Location:
> InMemoryFileIndex[file:/..., PartitionFilters: [], PushedFilers:
> [IsNotNull(day), EqualTo(day,20210702)], ReadSchema: struct<>, PushedFilters:
> [IsNotNull(day), EqualTo(day,20210702)]
> {code}
>
> Once the {{readDataSchema}} is not empty, it works correctly:
> {code:java}
> scala> spark.sql("SELECT header.tenant FROM parq WHERE day=20210702").explain
> == Physical Plan ==
> *(1) Project [header#51.tenant AS tenant#199]
> +- BatchScan[header#51, day#68] ParquetScan DataFilters: [], Format: parquet,
> Location: InMemoryFileIndex[file:/..., PartitionFilters: [isnotnull(day#68),
> (day#68 = 20210702)], PushedFilers: [IsNotNull(day), EqualTo(day,20210702)],
> ReadSchema: struct<header:struct<tenant:string>>, PushedFilters:
> [IsNotNull(day), EqualTo(day,20210702)]{code}
>
> In V1 this optimization is available:
> {code:java}
> scala> spark.sql("SELECT count(*) FROM parq WHERE day=20210702").explain
> == Physical Plan ==
> *(2) HashAggregate(keys=[], functions=[count(1)])
> +- Exchange SinglePartition, ENSURE_REQUIREMENTS, [id=#27]
> +- *(1) HashAggregate(keys=[], functions=[partial_count(1)])
> +- *(1) Project
> +- *(1) ColumnarToRow
> +- FileScan parquet [year#15,month#16,day#17,hour#18] Batched: true,
> DataFilters: [], Format: Parquet, Location: InMemoryFileIndex[file:/...,
> PartitionFilters: [isnotnull(day#17), (day#17 = 20210702)], PushedFilters:
> [], ReadSchema: struct<>{code}
> The examples use {{ParquetScan}}, but the problem happens for all File based
> V2 datasources.
> The fix for this issue feels very straight forward. In
> {{PruneFileSourcePartitions}} queries with an empty {{readDataSchema}} are
> explicitly excluded from being pushed down:
> {code:java}
> if filters.nonEmpty && scan.readDataSchema.nonEmpty =>{code}
> Removing that condition seems to fix the issue however, this might be too
> naive.
> I am making a PR with tests where this change can be discussed.
--
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
