[
https://issues.apache.org/jira/browse/HADOOP-1189?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Raghu Angadi updated HADOOP-1189:
---------------------------------
Status: Open (was: Patch Available)
I think we found the problem. Consider the following 'du .' out put:
Filesystem 1K-blocks Used Available Use% Mounted on
/dev/sda3 950877384 148532012 792685008 16% /export/workspace
'Capacity - Used' is 80234537 and available is 792685008. A little bit of
mismatch is expected. But this could be very large. On one node, we saw
'Capacity - Used' was 21GB and available was 0. In our code we use 'Capacity -
Used' instead of MIN(Capacity - Used, Available).
I will submit a new patch.
> Still seeing some unexpected 'No space left on device' exceptions
> -----------------------------------------------------------------
>
> Key: HADOOP-1189
> URL: https://issues.apache.org/jira/browse/HADOOP-1189
> Project: Hadoop
> Issue Type: Bug
> Components: dfs
> Affects Versions: 0.12.2
> Reporter: Raghu Angadi
> Assigned To: Raghu Angadi
> Fix For: 0.13.0
>
> Attachments: HADOOP-1189.patch
>
>
> One of the datanodes has one full partition (disk) out of four. Expected
> behaviour is that datanode should skip this partition and use only the other
> three. HADOOP-990 fixed some bugs related to this. It seems to work ok but
> some exceptions are still seeping through. In one case there 33 of these out
> 1200+ blocks written to this node. Not sure what caused this. I will submit a
> patch to the prints a more useful message throw the original exception.
> Two unlikely reasons I can think of are 2% reserve space (8GB in this case)
> is not enough or client some how still says block size is zero in some cases.
> Better error message should help here.
> If you see small number of these exceptions compared to number of blocks
> written, for now you don't need change anything.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
