[
https://issues.apache.org/jira/browse/OMID-131?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16812781#comment-16812781
]
Lars Hofhansl commented on OMID-131:
------------------------------------
Yeah not for 1.0.1.
I also want to make we're talking about the same thing. :)
Currently Omid writes a shadow-cell per data-cell. What we had discussed in the
past was to write a shadow-cell per row instead.
That would (1) keep the - already considerable - storage overhead of HBase
down, and (2) reduce stuff to be buffered at the client, (3) increase the
scanning speed of transactional tables. For wide tables this would be
sigifinicant in all those dimensions. Since all cells of a row are guaranteed
to be in the same region and HBase now has family (specific-)version delete
markers, this should be doable.
Is that what we're talking about? Or is this something else?
Note that #2 is dramatically improved with PHOENIX-5090 - if you can, please
have a look at my patch there.
> Enhance row level conlict detection to reduce client memory footprint.
> ----------------------------------------------------------------------
>
> Key: OMID-131
> URL: https://issues.apache.org/jira/browse/OMID-131
> Project: Apache Omid
> Issue Type: Improvement
> Affects Versions: 1.0.1
> Reporter: Yonatan Gottesman
> Assignee: Yonatan Gottesman
> Priority: Major
> Time Spent: 20m
> Remaining Estimate: 0h
>
> Currently even if omid is working in row level conflict detection, all the
> cells are kept by the client until after commit for writing shadow cells or
> removing them after abort.
> Instead the client should store only the rows or only the family.
> If we choose to store only the row, the client memory footprint will be
> smaller, but if a region server wants to query a shadow cell for a cell it
> might have to query a different region.
> On the other hand client can store the famelies of a writeset. This will have
> a bigger footprint if we have lots of families but if a region queries a
> shadow cell it will be on the same region.
>
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
