Member since
01-12-2015
1
Post
1
Kudos Received
0
Solutions
01-12-2015
05:23 AM
1 Kudo
Hi, PROBLEM: We update our cluster from CDH 5.1 to 5.3. After the upgrade some of our queries fail with permission denied. This happens when the user who runs the query queries a table over data owned by another user. EXAMPLE: For example, we have a Flume agent writing logs in "/user/flume/logs" with user flume. Hive table logs is created over that directory (with CREATE EXTERNAL TABLE). When we run a read-only query (SELECT) as user bi we get the following error: FAILED: RuntimeException Cannot create staging directory 'hdfs://nameservice/user/flume/logs/.hive-staging_hive_2015-01-12_10-43-00_285_2638530316386815724-1': Permission denied: user=bi, access=WRITE, inode="/user/flume/logs":flume:flume:drwxr-xr-x This is because the job run with user bi tries to create directory ".hive-staging_hive_2015-01-12_10-43-00_285_2638530316386815724-1" in "/user/flume/logs" which is owned by user flume and it does not have the permission to do this. In my opinion a read-only user, such as bi, shouldn't be able to write in a production directory such as "/user/flume/*", but they should be able to read it (which includes running queries on it). This never happend in previous versions of CDH. If I run a similar query with user bi over data owned by the same user bi, everything works fine, except that those ".hive-staging_hive*" directories are still created in table's location and they contain the whole data retrieved by the query, wasting our HDFS space. These directories should be anyway temporary and they should be deleted after a while. SOLUTION: After doing some research it seems that Hive needs to use a staging directory, which is configured by hive.exec.stagingdir configuration property. Before the upgrade this property had no value. After the upgrade it has value ".hive-staging". We changed it to "/tmp/hive-staging" and now everything works fine. Please fix this default value, as other CDH users might encounter it. Additionally, it would be fine to have this property in Cloudera Manager web interface. To make it worked we needed to put it in safety valve. Best regards, Călin-Andrei Burloiu.
... View more
Labels: