[jira] [Commented] (HBASE-10242) client-side mvcc tracking in scanners

["Sergey Shelukhin (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HBASE-10242?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13857341#comment-13857341 ] 

Sergey Shelukhin commented on HBASE-10242:
------------------------------------------

Problem (talking to myself): splits and merges.
Actually quite a bit of a problem, because by default the scanner will lose consistency on these - it's a new region so new mvcc will be obtained; but the simple alternative (tracking them and reusing mvcc) is even worse, esp. for merge - for example, if the first region had much lower mvcc before merge because it came from different server, the second region data with its old high mvcc numbers could become invisible. 
Probably, when obtaining the first mvcc for the region, the scanner will have track it for key range, and re-use it across splits / separate requests for merges.
Almost makes you wonder if getting mvccs in advance is worth it for consistent scanner - that would be one request to a few RSes (that hold the requisite regions).


> client-side mvcc tracking in scanners
> -------------------------------------
>
>                 Key: HBASE-10242
>                 URL: https://issues.apache.org/jira/browse/HBASE-10242
>             Project: HBase
>          Issue Type: Sub-task
>          Components: HFile, regionserver, Scanners
>            Reporter: Sergey Shelukhin
>            Assignee: Sergey Shelukhin
>
> Scanners should be able to track mvcc read point and send it to server. This is a subtask, so server can use this mvcc as best it can, but doesn't actually have to guarantee it within the scope of this jira.



--
This message was sent by Atlassian JIRA
(v6.1.5#6160)