[
https://issues.apache.org/jira/browse/CASSANDRA-14910?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16701035#comment-16701035
]
Aleksey Yeschenko commented on CASSANDRA-14910:
-----------------------------------------------
Force-pushed an updated version that just avoids making any decisions regarding
skipping sstables based on static slice bounds - start or end. Can't rely on
them, can't trust them, they are as dodgy as our recording of min/max cell
names if static rows are involved.
> Don't skip entire sstables when reading backwards with mixed clustering
> column order
> ------------------------------------------------------------------------------------
>
> Key: CASSANDRA-14910
> URL: https://issues.apache.org/jira/browse/CASSANDRA-14910
> Project: Cassandra
> Issue Type: Bug
> Components: Local Write-Read Paths
> Reporter: Aleksey Yeschenko
> Assignee: Aleksey Yeschenko
> Priority: Major
> Fix For: 2.2.x
>
>
> In 2.x, if a table has clustering columns in {{DESC}} order, any SSTable that
> doesn’t have any static rows in it will be skipped altogether when iterated
> in reverse.
> This occurs due to the logic in {{ColumnSlice.compare()}} errorneusly sorting
> any empty {{ByteBuffer}} after non-empty values due to the way
> {{ReversedType}} operates. In case that empty {{ByteBuffer}} is coming from a
> static {{Composite}}, however, the logic breaks down. Static {{Composite}}
> components must *always* sort before any non-empty value, no matter the
> table’s comparator.
> 2.0, 2.1, and 2.2 are all affected. 3.0 isn’t, but only because we skip slice
> intersection logic entirely if static rows are present in the filter.
> Introduced by CASSANDRA-8502.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
