[
https://issues.apache.org/jira/browse/KAFKA-596?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Swapnil Ghike updated KAFKA-596:
--------------------------------
Attachment: kafka-596.patch
Yes, that's a pretty sharp observation. The current time based roll policy only
checks whether the segment was not appended for its lifetime. This patch has
three fixes:
1. Initial assignment of firstAppendTime in Logsegment, because the non-primary
constructor could potentially initialize the messageSet and set its size > 0.
(I don't know if this fix will affect any other jiras.)
2. In maybeRoll(), a new condition makes sure that roll() happens based on time
only if the messageset size > 0, thus different segments cannot have identical
starting offsets. It also makes sure that a new segment is not rolled if the
last segment is not appended with messages until now.
2. A segment is reborn at three places by setting its firstAppendTime to None
if the message set size is 0 -
i. Log.maybeRoll()
ii. Logsement.truncateTo()
iii. Log.markedDeletedWhile()
> LogSegment.firstAppendTime not reset after truncate to
> ------------------------------------------------------
>
> Key: KAFKA-596
> URL: https://issues.apache.org/jira/browse/KAFKA-596
> Project: Kafka
> Issue Type: Bug
> Components: core
> Affects Versions: 0.8
> Reporter: Jun Rao
> Assignee: Swapnil Ghike
> Labels: bugs
> Attachments: kafka-596.patch
>
> Original Estimate: 24h
> Remaining Estimate: 24h
>
> Currently, we don't reset LogSegment.firstAppendTime after the segment is
> truncated. What can happen is that we truncate the segment to size 0 and on
> next append, a new log segment with the same starting offset is rolled
> because the time-based rolling is triggered.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
