hive-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Selina Zhang (JIRA)" <>
Subject [jira] [Commented] (HIVE-6980) Drop table by using direct sql
Date Mon, 28 Apr 2014 21:41:15 GMT


Selina Zhang commented on HIVE-6980:

[~ashutoshc] Thanks! I have checked HIVE-6809. That patch would not help the use case of drop
table. In HIVE-6809, the main problem it is trying to solve is optimize the deletion of partition
data. Instead delete the partition path one by one, the new patch tried to delete the paths
in bulk. However, in drop table case, this optimization has been done(the method isSubdirectory()
check if the partition path is subdirectory of table path). 

The problem causes dropping table very slow is how we drop partitions. 
1.In client side, DDLTask asks server for each Partition object in order to check if the partition
is protected. So if we have 100K partitions, 100K Partition objects will be constructed and
passing along from the server to client (in a batch way, so won't be OOM here). 
2.After the check, client sends drop table request to the server. And HiveMetaStore begins
the fun again. It retrieves each Partition object again and check the data path. Then send
the partition names to ObjectStore and ask for deletion (in batch mode).
3. The ObjectStore got a list of partition names. With the fix of HIVE-6252, the DataNucleus
will do the real work. It cleans the related metadata tables for each partition. I have turned
on the query log for MySQL. For dropping a table with 6 partitions, there were 72 DELETEs,

> Drop table by using direct sql
> ------------------------------
>                 Key: HIVE-6980
>                 URL:
>             Project: Hive
>          Issue Type: Improvement
>          Components: Metastore
>    Affects Versions: 0.12.0
>            Reporter: Selina Zhang
>            Assignee: Selina Zhang
> Dropping table which has lots of partitions is slow. Even after applying the patch of
HIVE-6265, the drop table still takes hours (100K+ partitions). 
> The fixes come with two parts:
> 1. use directSQL to query the partitions protect mode;
> the current implementation needs to transfer the Partition object to client and check
the protect mode for each partition. I'd like to move this part of logic to metastore. The
check will be done by direct sql (if direct sql is disabled, execute the same logic in the
> 2. use directSQL to drop partitions for table;
> there maybe two solutions here:
> 1. add "DELETE CASCADE" in the schema. In this way we only need to delete entries from
partitions table use direct sql. May need to change datanucleus.deletionPolicy = DataNucleus.

> 2. clean up the dependent tables by issue DELETE statement. This also needs to turn on
> Both of above solutions should be able to fix the problem. The DELETE CASCADE has to
change schemas and prepare upgrade scripts. The second solutions added maintenance cost if
new tables added in the future releases.
> Please advice. 

This message was sent by Atlassian JIRA

View raw message