When Spark SQL infers a schema for a DataFrame, it will take the union of all field types present in the structured source data (e.g. an RDD of JSON data). When the source data for a row doesn't define a particular field on the DataFrame's schema, a null value will simply be assumed for this field. This workflow makes it very easy to construct tables and query over a set of structured data with a nonuniform schema. However, this behavior is not consistent in some cases when dealing with Parquet files and an external table managed by an external Hive metastore.
In our particular usecase, we use Spark Streaming to parse and transform our input data and then apply a window function to save an arbitrary-sized batch of data as a Parquet file, which itself will be added as a partition to an external Hive table via an "ALTER TABLE... ADD PARTITION..." statement. Since our input data is nonuniform, it is expected that not every partition batch will contain every field present in the table's schema obtained from the Hive metastore. As such, we expect that the schema of some of our Parquet files may not contain the same set fields present in the full metastore schema.
In such cases, it seems natural that Spark SQL would simply assume null values for any missing fields in the partition's Parquet file, assuming these fields are specified as nullable by the metastore schema. This is not the case in the current implementation of ParquetRelation2. The mergeMetastoreParquetSchema() method used to reconcile differences between a Parquet file's schema and a schema retrieved from the Hive metastore will raise an exception if the Parquet file doesn't match the same set of fields specified by the metastore.
I propose altering this implementation in order to allow for any missing metastore fields marked as nullable to be merged in to the Parquet file's schema before continuing with the checks present in mergeMetastoreParquetSchema().
Classifying this as a bug as it exposes inconsistent behavior, IMHO. If you feel this should be an improvement or new feature instead, please feel free to reclassify this issue.