[
https://issues.apache.org/jira/browse/HBASE-9904?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13816300#comment-13816300
]
Jimmy Xiang commented on HBASE-9904:
------------------------------------
In trunk, we use call sequence number to prevent such thing. Due to retry, if
the client is not on the same page with the region server, it will get an
exception instead of next batch of data. This test works fine with trunk. So I
don't think this is an issue with trunk any more.
> Solve skipping data in HTable scans
> -----------------------------------
>
> Key: HBASE-9904
> URL: https://issues.apache.org/jira/browse/HBASE-9904
> Project: HBase
> Issue Type: Bug
> Components: Client
> Affects Versions: 0.89-fb
> Reporter: Manukranth Kolloju
> Priority: Critical
> Fix For: 0.89-fb, 0.98.0
>
> Attachments: scan.diff
>
>
> The HTable client cannot retry a scan operation in the
> getRegionServerWithRetries code path.
> This will result in the client missing data. This can be worked around using
> hbase.client.retries.number to 1.
> The whole problem is that Callable knows nothing about retries and the
> protocol it dances to as well doesn't support retires.
> This fix will keep Callable protocol (ugly thing worth merciless refactoring)
> intact but will change
> ScannerCallable to anticipate retries. What we want is to make failed
> operations to be identities for outside world:
> N1 , N2 , F3 , N3 , F4 , F4 , N4 ... = N1 , N2 , N3 , N4 ...
> where Nk are successful operation and Fk are failed operations.
--
This message was sent by Atlassian JIRA
(v6.1#6144)
