[
https://issues.apache.org/jira/browse/SPARK-58711?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
ASF GitHub Bot updated SPARK-58711:
-----------------------------------
Labels: pull-request-available (was: )
> Use StructType.getFieldIndex in Parquet and ORC aggregate push down
> -------------------------------------------------------------------
>
> Key: SPARK-58711
> URL: https://issues.apache.org/jira/browse/SPARK-58711
> Project: Spark
> Issue Type: Improvement
> Components: SQL
> Affects Versions: 4.3.0
> Reporter: Uroš Bojanić
> Priority: Major
> Labels: pull-request-available
>
> The aggregate push down paths look up a column position with
> dataSchema.fieldNames.indexOf(name). fieldNames is a def that allocates a
> fresh Array on every call, and the Parquet variant additionally converts it
> to a List, so each lookup allocates twice and then scans linearly.
> StructType.getFieldIndex reads a cached name to index map instead.
> This switches the three Parquet sites and the one ORC site to getFieldIndex,
> and replaces a partitionSchema.fields.map(_.name).toSet.contains check, which
> built a set per call, with getFieldIndex(...).isDefined.
> This is a readability and allocation cleanup rather than a measurable
> speedup: aggregate push down is off by default and the surrounding work reads
> Parquet or ORC footer statistics, which dominates.
> Behavior is unchanged. getOrElse(-1) matches the not-found contract of
> indexOf, and the exact-match variant is used rather than the case-insensitive
> one. The two forms would differ only for a schema with duplicate column
> names, where indexOf returns the first match and a name keyed map the last;
> FileTable rejects duplicate column names in both schemas before a scan is
> built, so that case is unreachable.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
