[
https://issues.apache.org/jira/browse/SPARK-21841?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16145325#comment-16145325
]
Wenchen Fan commented on SPARK-21841:
-------------------------------------
This is above the data source layer. When you create a data source table, i.e.
{{t1.write.saveAsTable('mydb.t1')}} , spark will keep the table schema in table
properties and ignore schema updates from hive. When you create a hive serde
table, i.e., {{t1.write.format("hive").saveAsTable('mydb.t1')}} , then it can
work as you expect, because hive serde table will always respect table schema
from hive metastore.
I think the confusing part is, for `DataFrameWriter`, by default it creates
parquet data source table, while for `CREATE TABLE` statement, by default it
creates text hive serde table. This is due to some historical reasons, but we
may need to document it more explicitly.
> Spark SQL doesn't pick up column added in hive when table created with
> saveAsTable
> ----------------------------------------------------------------------------------
>
> Key: SPARK-21841
> URL: https://issues.apache.org/jira/browse/SPARK-21841
> Project: Spark
> Issue Type: Bug
> Components: SQL
> Affects Versions: 2.1.0, 2.2.0
> Reporter: Thomas Graves
>
> If you create a table in Spark sql but then you modify the table in hive to
> add a column, spark sql doesn't pick up the new column.
> Basic example:
> {code}
> t1 = spark.sql("select ip_address from mydb.test_table limit 1")
> t1.show()
> +------------+
> | ip_address|
> +------------+
> |1.30.25.5|
> +------------+
> t1.write.saveAsTable('mydb.t1')
> In Hive:
> alter table mydb.t1 add columns (bcookie string)
> t1 = spark.table("mydb.t1")
> t1.show()
> +------------+
> | ip_address|
> +------------+
> |1.30.25.5|
> +------------+
> {code}
> It looks like its because spark sql is picking up the schema from
> spark.sql.sources.schema.part.0 rather then from hive.
> Interestingly enough it appears that if you create the table differently like:
> spark.sql("create table mydb.t1 select ip_address from mydb.test_table limit
> 1")
> Run your alter table on mydb.t1
> val t1 = spark.table("mydb.t1")
> Then it works properly.
> It looks like the difference is when it doesn't work
> spark.sql.sources.provider=parquet is set.
> Its doing this from createDataSourceTable where provider is parquet.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
