[
https://issues.apache.org/jira/browse/HADOOP-1134?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12485339
]
Sameer Paranjpye commented on HADOOP-1134:
------------------------------------------
> Why wouldn't the map tasks run on a node where the block is local? The
> checksum data would need to be read over the network, but checksums are 1%
> the size of data, and
> we typically assume that net reads from a random node are 10x slower than
> local disk reads, so the checksum network i/o should only add 10% to the cost
> of reading the
> block, right?
Yes, it could be done that way, if each split were a set of block instances on
a node. The client would need a way to go from a block id to a .crc file via an
extension of the Namenode API. The difficulty there is in determining the set
of validated files from the set of validated blocks and so knowing which .crc
files can be deleted. Of course, all the .crc files could be deleted at the end.
The way I was thinking about it was to have each split be a file or a set of
files, it would be hard to schedule local to all the blocks in that case. This
requires practically no API changes, there already exists a API to report
corrupt blocks. Once a file is validated the .crc file would be deleted by the
client. The set of .crc files remaining at the end tells you exactly which data
is suspect. This feels very clean, but doesn't do such a great job of ensuring
data locality.
> Block level CRCs in HDFS
> ------------------------
>
> Key: HADOOP-1134
> URL: https://issues.apache.org/jira/browse/HADOOP-1134
> Project: Hadoop
> Issue Type: New Feature
> Components: dfs
> Reporter: Raghu Angadi
> Assigned To: Raghu Angadi
>
> Currently CRCs are handled at FileSystem level and are transparent to core
> HDFS. See recent improvement HADOOP-928 ( that can add checksums to a given
> filesystem ) regd more about it. Though this served us well there a few
> disadvantages :
> 1) This doubles namespace in HDFS ( or other filesystem implementations ). In
> many cases, it nearly doubles the number of blocks. Taking namenode out of
> CRCs would nearly double namespace performance both in terms of CPU and
> memory.
> 2) Since CRCs are transparent to HDFS, it can not actively detect corrupted
> blocks. With block level CRCs, Datanode can periodically verify the checksums
> and report corruptions to namnode such that name replicas can be created.
> We propose to have CRCs maintained for all HDFS data in much the same way as
> in GFS. I will update the jira with detailed requirements and design. This
> will include same guarantees provided by current implementation and will
> include a upgrade of current data.
>
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
