[
https://issues.apache.org/jira/browse/HADOOP-1134?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12482872
]
Raghu Angadi commented on HADOOP-1134:
--------------------------------------
Correctness is a big advantage of using existing CRCs. Still, we could choose
to 64k checksums during upgrade, which implies reading of of all the blocks and
associated increase in downtime. Whether to recreate checksums or not depends
on how far off is 512 byte CRCs is from our (gu)estimated optimal value.
Another option to get more trust after to upgrade is to compare checksums of
replicas and choose the majority. For current design we can assume we compare
with the old CRCs.
If a client asks for 2k data, we are saying we should send 64k (or 128k) where
2k is located to client. Why not send only 2k and send newly calculated CRC for
2k? Datanode still verifies all the 64k blocks involved in the read but
recalculates. One argument against this is that this is a weaker guarantee than
sending full blocks for verification. But datanode will calculate the new CRC
right next to where the on disk CRC is verified thus minimizing some other
corruptions. I feel this compromise is probably worth it. But I guess many will
disagree. Requiring whole blocks will also increase overhead when support
appends in future.
How do we benchmark for good CRC-chunk size? It heavily depends on work load. I
will find more about a typical MapReduce load.
> 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.
