[ 
https://issues.apache.org/jira/browse/HBASE-4344?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13113387#comment-13113387
 ] 

Amitanand Aiyer commented on HBASE-4344:
----------------------------------------

@stack: Yes. I agree, this is not really cleaned up. Wanted to put it out as v1 
to get feedback on the approach.

One of the reasons why I went with adding the memstoreTS to HFile was indeed to 
minimize changes. Although, after
getting my feet wet, I did also consider moving along the direction of 
modifying the byte buffer itself to include
the memstoreTS. 

I felt like this would be simpler. The way I like to see it, memstoreTS is some 
"system level timestamp" that is 
somewhat transparant to the client/users. Wanted to keep it that way, as the 
key-values that are in memstore keep
the timestamp separate.

I agree about the fact that we are exposing the key-value internals to the 
HFile writer/readers. Perhaps, we should
get rid of it. Will be happy to do that, as a separate diff, once we get the 
core issue fixed.

also working on a diff to see if we can use the hlog sequence id as the 
memstoreTS ... will put it out separately.

> Persist memstoreTS to disk
> --------------------------
>
>                 Key: HBASE-4344
>                 URL: https://issues.apache.org/jira/browse/HBASE-4344
>             Project: HBase
>          Issue Type: Sub-task
>            Reporter: Amitanand Aiyer
>            Assignee: Amitanand Aiyer
>             Fix For: 0.89.20100924
>
>         Attachments: 4344-v2.txt, 4344-v4.txt, 4344-v5.txt, 4344-v6.txt, 
> 4344-v7.txt, 4344-v8.txt, patch-2
>
>
> Atomicity can be achieved in two ways -- (i) by using  a multiversion 
> concurrency system (MVCC), or (ii) by ensuring that "new" writes do not 
> complete, until the "old" reads complete.
> Currently, Memstore uses something along the lines of MVCC (called RWCC for 
> read-write-consistency-control). But, this mechanism is not incorporated for 
> the key-values written to the disk, as they do not include the memstore TS.
> Let us make the two approaches be similar, by persisting the memstoreTS along 
> with the key-value when it is written to the disk.

--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira

        

Reply via email to