[
https://issues.apache.org/jira/browse/IMPALA-7184?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16804227#comment-16804227
]
Thomas Tauber-Marshall commented on IMPALA-7184:
------------------------------------------------
[~andrew.wong] I played around with this and it doesn't work, at least not as
Impala expects it to work:
- I create a Kudu table, insert some stuff into it, scan it back at
READ_YOUR_WRITES. Everything works as expected.
- I wait greater than 'tablet_history_max_age_sec' and attempt to scan it again
at READ_YOUR_WRITES (using either the same KuduClient or a new one). This
results in an error of the form 'Snapshot timestamp is earlier than the ancient
history mark...'
I can avoid the error if I interact with the table in some other way (eg.
performing a scan at READ_LATEST or do an ALTER) and then scan it at
READ_YOUR_WRITES in less than 'tablet_history_max_age_sec'
I can avoid the error if I call SetLatestObservedTimestamp() to something that
is more recent than 'tablet_history_max_age_sec', but currently Impala only
calls SetLatestObservedTimestamp() if you're in a session in which there was a
previous DML operation (in which case its set to the return of
GetLatestObservedTimestamp() just after the dml has been flushed for the last
time). Is the expectation that Impala should always be calling
SetLatestObservedTimestamp() before scans?
> Support Kudu's READ_YOUR_WRITES scan mode
> -----------------------------------------
>
> Key: IMPALA-7184
> URL: https://issues.apache.org/jira/browse/IMPALA-7184
> Project: IMPALA
> Issue Type: Improvement
> Reporter: Thomas Tauber-Marshall
> Assignee: Thomas Tauber-Marshall
> Priority: Major
> Labels: kudu
>
> Kudu recently added a new scan mode called READ_YOUR_WRITES which provides
> better consistency guarantees than READ_LATEST or READ_AT_SNAPSHOT, the
> options currently supported by Impala.
> Unfortunately, READ_YOUR_WRITES is currently affected by a bug that makes it
> unusable by Impala (KUDU-2233). Once this is fixed, we should add support for
> it, and consider either setting it as the default, or at least using it in
> tests, see the discussion in https://gerrit.cloudera.org/#/c/10503/
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-all-unsubscr...@impala.apache.org
For additional commands, e-mail: issues-all-h...@impala.apache.org
