[
https://issues.apache.org/jira/browse/PHOENIX-4119?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16141291#comment-16141291
]
Mark Christiaens commented on PHOENIX-4119:
-------------------------------------------
Badly synchronous clocks was the way that we ran into this issue. Properly
enabling NTP helped suppress the error but clearly, perfectly synchronized
clocks are physically impossible so the issue could still pop up.
> Upsert uses timestamp of server holding table metadata
> ------------------------------------------------------
>
> Key: PHOENIX-4119
> URL: https://issues.apache.org/jira/browse/PHOENIX-4119
> Project: Phoenix
> Issue Type: Bug
> Affects Versions: 4.8.0
> Reporter: Mark Christiaens
>
> When doing an upsert, I noticed that the resulting put-command to HBase uses
> a timestamp obtained from the server managing the corresponding table's
> metadata (cfr.
> https://issues.apache.org/jira/browse/PHOENIX-1674?focusedCommentId=14349982&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-14349982).
>
> Now, if the row that is updated resides on a second server, that second
> server could have some clock skew. After the upsert, if you fetch that
> updated row without specifying an explicit query timestamp, then you might
> not see the effects of the upsert until the second server's clock has caught
> up with the first server's clock.
> I think that the desired behavior is that the puts performed occur with a
> timestamp derived from the region server where their rows are stored. (Of
> course, that still leaves the problem of region migration from one server to
> the next but it's a step in the right direction).
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
