[
https://issues.apache.org/jira/browse/CASSANDRA-13282?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15892849#comment-15892849
]
Jeff Jirsa commented on CASSANDRA-13282:
----------------------------------------
Attaching a drawing for whichever reviewer wants this ticket - drawing created
while discussing this offline because it's somewhat nuanced apparently.
Basically when we allocate in {{sync()}}, if we're at the end of a file, we
return -1, and then the end marker for the segment gets set to the end of the
file. Therefore within the while loop as we replay an individual sync section,
we can get to a point where we throw trying to read an int from the unused tail
of the section.
> Commitlog replay may fail if last mutation is within 4 bytes of end of segment
> ------------------------------------------------------------------------------
>
> Key: CASSANDRA-13282
> URL: https://issues.apache.org/jira/browse/CASSANDRA-13282
> Project: Cassandra
> Issue Type: Bug
> Components: Core
> Reporter: Jeff Jirsa
> Assignee: Jeff Jirsa
> Fix For: 2.2.x, 3.0.x, 3.11.x, 4.x
>
> Attachments: whiteboard.png
>
>
> Following CASSANDRA-9749 , stricter correctness checks on commitlog replay
> can incorrectly detect "corrupt segments" and stop commitlog replay (and
> potentially stop cassandra, depending on the configured policy). In
> {{CommitlogReplayer#replaySyncSection}} we try to read a 4 byte int
> {{serializedSize}}, and if it's 0 (which will happen due to zeroing when the
> segment was created), we continue on to the next segment. However, it appears
> that if a mutation is sized such that it ends with 1, 2, or 3 bytes remaining
> in the segment, we'll pass the {{isEOF}} on the while loop but fail to read
> the {{serializedSize}} int, and fail.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
