[
https://issues.apache.org/jira/browse/HDFS-4885?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13783405#comment-13783405
]
Luke Lu commented on HDFS-4885:
-------------------------------
I posted the comment before refreshing to see the v2 patch, which is getting
closer. I'd use something like {{hasMisplacement}} to check misplacement in
fsck instead of {{getMisReplicatedNum}}, as the latter might not be easy
calculate for certain BBPs. {{BlockPlacementStatus#getDescription}} is probably
better than either {{toString}} or {{getMisReplicatedString}}.
One the second thought, I wonder if we should simply pass down the
{{HdfsFileStatus}} instead of just a string to {{verifyBlockPlacement}} and get
rid of the constraints creation in fsck, as the file status will contain the
needed info to construct/enforce proper constraints in the BPP impls. So the
API would be:
{code}
public BlockPlacementStatus verifyBlockPlacement(String srcParent,
HdfsFileStatus srcFile, LocatedBlock blk);
{code}
The advantage of this is that if we have (extended) attributes to implement
various placement grouping policies in the future, no API change is necessary.
We should probably add a new API to calculate/reset the accumulative status of
a BBP instance, so a proper summary can be printed. This is also useful to
verify group (anti-)affinity, which cannot be determined on a block by block
basis.
{code}
public void resetStatus();
public BlockPlacementStatus getStatus();
{code}
Thoughts?
> Update verifyBlockPlacement() API in BlockPlacementPolicy
> ---------------------------------------------------------
>
> Key: HDFS-4885
> URL: https://issues.apache.org/jira/browse/HDFS-4885
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Reporter: Junping Du
> Assignee: Junping Du
> Labels: BlockPlacementPolicy
> Attachments: HDFS-4885.patch, HDFS-4885-v2.patch
>
>
> verifyBlockPlacement() has unused parameter -srcPath as its responsibility
> just verify single block rather than files under a specific path. Also the
> return value (int) does not make sense as the violation of block placement
> has other case than number of racks, so boolean value should be better.
--
This message was sent by Atlassian JIRA
(v6.1#6144)
