[
https://issues.apache.org/jira/browse/SPARK-60558?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Gengliang Wang reassigned SPARK-60558:
--------------------------------------
Assignee: Yan Yan
> Expose subquery planning context to DataSource V2 scan builders
> ---------------------------------------------------------------
>
> Key: SPARK-60558
> URL: https://issues.apache.org/jira/browse/SPARK-60558
> Project: Spark
> Issue Type: Improvement
> Components: SQL
> Affects Versions: 5.0.0
> Reporter: Yan Yan
> Assignee: Yan Yan
> Priority: Major
> Labels: pull-request-available
>
> DataSource V2 connectors can perform expensive statistics computation while
> creating and planning a scan. Spark currently does not tell an ordinary batch
> table reader whether its scan builder is being created within a subquery, so
> the connector cannot use that scope to decide how much statistics work to
> perform.
> Add a ScanContext argument to a new SupportsRead.newScanBuilder overload and
> provide the current subquery planning scope when Spark constructs the
> builder. The default overload delegates to the existing options-only method
> for compatibility with existing connectors. A connector can use the context
> to avoid expensive statistics work while planning a subquery, without
> changing the rows returned by its scan.
> The context is advisory and describes the scope at builder creation. Later
> optimizer rewrites may move or reuse the scan in another scope; a newly
> constructed builder receives a new context for its destination scope. This is
> not a guarantee about the final query shape.
> This change covers ordinary batch reads. Context overloads for row-level
> operations and changelog reads can be considered separately when needed.
> PR: https://github.com/apache/spark/pull/59662
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
