Major Description
Although the SparkSqlSerializer class extends the KryoSerializer in core, it's overridden newKryo() does not call super.newKryo(). This results in inconsistent serializer behaviors depending on whether a KryoSerializer instance or a SparkSqlSerializer instance is used. This may also be related to the TODO in KryoResourcePool, which uses KryoSerializer instead of SparkSqlSerializer due to yet-to-be-investigated test failures.
An example of the divergence in behavior: The Exchange operator creates a new SparkSqlSerializer instance (with an empty conf; another issue) when it is constructed, whereas the GENERIC ColumnType pulls a KryoSerializer out of the resource pool (see above). The result is that the serialized in-memory columns are created using the user provided serializers / registrators, while serialization during exchange does not.
