Details
Description
Metastore needs removing unused column descriptors when drop/add partitions or tables. For query the unused column descriptor, the current implementation utilizes datanuleus' range function, which basically equals LIMIT syntax. However, Oracle does not support LIMIT, the query is converted as
SQL> SELECT * FROM (SELECT subq.*,ROWNUM rn FROM (SELECT
'org.apache.hadoop.hive.metastore.model.MStorageDescriptor' AS
NUCLEUS_TYPE,A0.INPUT_FORMAT,A0.IS_COMPRESSED,A0.IS_STOREDASSUBDIRECTORIES,A0.LOCATION,
A0.NUM_BUCKETS,A0.OUTPUT_FORMAT,A0.SD_ID FROM drhcat.SDS A0
WHERE A0.CD_ID = ? ) subq ) WHERE rn <= 1;
Given that CD_ID is not very selective, this query may have to access large amount of rows (depends how many partitions the table has, millions of rows in our case). Metastore may become unresponsive because of this.
Since Metastore only needs to know if the specific CD_ID is referenced in SDS table and does not need access the whole row. We can use
select count(1) from SDS where SDS.CD_ID=?
CD_ID is index column, the above query will do range scan for index, which is faster.
For other DBs support LIMIT syntax such as MySQL, this problem does not exist. However, the new query does not hurt.
Attachments
Attachments
Issue Links
- breaks
-
HIVE-21075 Metastore: Drop partition performance downgrade with Postgres DB
- Closed