Description
The given example is if datanode disk definitions but should be applicable to all configuration where a list of disks are provided.
I have defined multiple local disks defined for a datanode:
<property>
<name>dfs.data.dir</name>
<value>/data/01/dfs/dn,/data/02/dfs/dn,/data/03/dfs/dn,/data/04/dfs/dn,/data/05/dfs/dn,/data/06/dfs/dn</value>
<final>true</final>
</property>
When one of those disks breaks and is unmounted then the mountpoint (such as /data/03 in this example) becomes a regular directory which doesn't have the valid permissions and possible directory structure Hadoop is expecting.
When this situation happens, the datanode fails to restart because of this while actually we have enough disks in an OK state to proceed. The only way around this is to alter the configuration and omit that specific disk configuration.
To my opinion, It would be more practical to let Hadoop daemons start when at least 1 disks/partition in the provided list is in a usable state. This prevents having to roll out custom configurations for systems which have temporarily a disk (and therefor directory layout) missing. This might also be configurable that at least X partitions out of he available ones are in OK state.
Attachments
Issue Links
- duplicates
-
HDFS-1592 Datanode startup doesn't honor volumes.tolerated
-
- Closed
-