Gengliang Wang created SPARK-27356:
--------------------------------------
Summary: File source V2: return actual schema in method
`FileScan.readSchema`
Key: SPARK-27356
URL: https://issues.apache.org/jira/browse/SPARK-27356
Project: Spark
Issue Type: Improvement
Components: SQL
Affects Versions: 3.0.0
Reporter: Gengliang Wang
The method `Scan.readSchema` returns the actual schema of this data source
scan.
In the current file source V2 framework, the schema is not returned correctly.
The result should be `readDataSchema + partitionSchema`, while the framework
returns the schema which is pushed down in
`SupportsPushDownRequiredColumns.requiredSchema`.
This is normally OK. But if there are overlap columns between `dataSchema` and
`partitionSchema`, the result of row-base scan will be wrong. The actual schema
should be
`dataSchema - overlapSchema + partitionSchema`, which is different from from
the pushed down `requiredSchema`.
This PR is to:
1. Bug fix: fix the corner case that `dataSchema` overlaps with
`partitionSchema`.
2. Improvement: Prune partition column values if part of the partition columns
are not required.
3. Behavior change: To make it simple, the schema of `FileTable` is `dataSchema
- overlapSchema + partitionSchema`, instead of mixing data schema and
partitionSchema (see `PartitioningUtils.mergeDataAndPartitionSchema`)
For example, the data schema is [a,b,c], the partition schema is [b,d],
In V1, the schema of `HadoopFsRelation` is [a, b, c, d]
in File source V2 , the schema of `FileTable` is [a, c, b, d]
Putting all the partition columns to the end of table schema is more
reasonable. Also, when there is `select *` operation and there is no schema
pruning, the schema of `FileTable` and `FileScan` still matches.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
