[ 
https://issues.apache.org/jira/browse/HDFS-12283?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16128428#comment-16128428
 ] 

Yuanbo Liu commented on HDFS-12283:
-----------------------------------

[~cheersyang] Thanks a lot for your review.
Agree with the most part of your comments and I've addressed them in v2 patch.
{quote}
line 110: this might have problem.. e.g you have listed tx 1,3,5 (2,4 has been 
committed), then 4 won't be a valid start key anymore.
{quote}
It's quite a tricky part because committing transactions could happen in any 
time. So we decide to use a filter instead of  a startKey to avoid repeatedly 
scanning the same records. The whole idea is based on the sequentially 
increasing txid,  we record the last read txid, scan the db from the very 
beginning, and get the transactions whose id are larger than the last read 
txid, so that the invalid start key issue could be addressed.

> Ozone: DeleteKey-5: Implement SCM DeletedBlockLog
> -------------------------------------------------
>
>                 Key: HDFS-12283
>                 URL: https://issues.apache.org/jira/browse/HDFS-12283
>             Project: Hadoop HDFS
>          Issue Type: Sub-task
>          Components: ozone, scm
>            Reporter: Weiwei Yang
>            Assignee: Yuanbo Liu
>         Attachments: HDFS-12283.001.patch, HDFS-12283-HDFS-7240.001.patch, 
> HDFS-12283-HDFS-7240.002.patch
>
>
> The DeletedBlockLog is a persisted log in SCM to keep tracking container 
> blocks which are under deletion. It maintains info about under-deletion 
> container blocks that notified by KSM, and the state how it is processed. We 
> can use RocksDB to implement the 1st version of the log, the schema looks like
> ||TxID||ContainerName||Block List||ProcessedCount||
> |0|c1|b1,b2,b3|0|
> |1|c2|b1|3|
> |2|c2|b2, b3|-1|
> Some explanations
> # TxID is an incremental long value transaction ID for ONE container and 
> multiple blocks
> # Container name is the name of the container
> # Block list is a list of block IDs
> # ProcessedCount is the number of times SCM has sent this record to datanode, 
> it represents the "state" of the transaction, it is in range of \[-1, 5\], -1 
> means the transaction eventually failed after some retries, 5 is the max 
> number times of retries.
> We need to define {{DeletedBlockLog}} as an interface and implement this with 
> RocksDB {{MetadataStore}} as the first version.



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)

---------------------------------------------------------------------
To unsubscribe, e-mail: hdfs-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: hdfs-issues-h...@hadoop.apache.org

Reply via email to