[
https://issues.apache.org/jira/browse/HDFS-7354?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14288556#comment-14288556
]
Zhe Zhang commented on HDFS-7354:
---------------------------------
[~szetszwo] Good point, under the striping layout, we don't group multiple
files in the same block group. So each parity block can indeed be treated as
part of the file.
Parity blocks still need special treatment though, because they should be
recovered with lower priority than data blocks. I'll update the JIRA
description.
This JIRA was actually created based on the [old design |
https://issues.apache.org/jira/secure/attachment/12677810/HDFSErasureCodingDesign-20141028.pdf]
with non-striping / contiguous layout, where parity blocks could be orphans.
We chose not to implement it in the initial phase because it complicates file
deletions. But I don't think it's completely off the table. We might revisit it
in the future to work with applications which have strict locality
requirements.
> Support parity blocks in block management
> -----------------------------------------
>
> Key: HDFS-7354
> URL: https://issues.apache.org/jira/browse/HDFS-7354
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Reporter: Zhe Zhang
> Assignee: Zhe Zhang
>
> In the current block management system, each block is associated with a file.
> "Orphan" blocks are considered corrupt and will be removed.
> In this JIRA we extend {{Block}} with a binary flag denoting whether it is a
> parity block ({{isParity}}). Parity blocks are created, stored, and reported
> the same way as raw ones. They have regular block IDs which are unrelated to
> those of the raw blocks in the same group; their replicas (normally only 1)
> are stored in RBW and finalized directories on the DataNode depending on the
> stage; they are also included in block reports. The only distinction of a
> parity block is the lack of file affiliation. The block management system
> will be aware of parity blocks and will _not_ try to remove them.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
