[SPARK-19455] Add option for case-insensitive Parquet field resolution

• Type: Improvement
• Status: Closed
• Priority: Major
• Resolution: Duplicate
• Affects Version/s: 2.1.0
• Component/s: SQL
• Assignee: Unassigned
• Reporter: Adam Budde
• Created: 03/Feb/17 22:32
• Updated: 15/Feb/17 19:41
• Resolved: 15/Feb/17 19:41

Description

SPARK-16980 removed the schema inferrence from the HiveMetastoreCatalog class when converting a MetastoreRelation to a LoigcalRelation (HadoopFsRelation, in this case) in favor of simply using the schema returend by the metastore. This results in an optimization as the underlying file status no longer need to be resolved until after the partition pruning step, reducing the number of files to be touched significantly in some cases. The downside is that the data schema used may no longer match the underlying file schema for case-sensitive formats such as Parquet.

This change initially included a patch to ParquetReadSupport that attempted to remedy this conflict by using a case-insentive fallback mapping when resolving field names during the schema clipping step. SPARK-1833 later removed this patch after SPARK-17183 added support for embedding a case-sensitive schema as a Hive Metastore table property. AFAIK the assumption here was that the data schema obtained from the Metastore table property will be case sensitive and should match the Parquet schema exactly.

The problem arises when dealing with Parquet-backed tables for which this schema has not been embedded as a table attributes and for which the underlying files contain case-sensitive field names. This will happen for any Hive table that was not created by Spark or created by a version prior to 2.1.0. We've seen Spark SQL return no results for any query containing a case-sensitive field name for such tables.

The change we're proposing is to introduce a configuration parameter that will re-enable case-insensitive field name resolution in ParquetReadSupport. This option will also disable filter push-down for Parquet, as the filter predicate constructed by Spark SQL contains the case-insensitive field names which Parquet will return 0 records for when filtering against a case-sensitive column name. I was hoping to find a way to construct the filter on-the-fly in ParquetReadSupport but Parquet doesn't propegate the Configuration object passed to this class to the underlying InternalParquetRecordReader class.

Issue Links

• links to: [Github] Pull Request #16797 (budde)