[
https://issues.apache.org/jira/browse/CASSANDRA-20190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17936447#comment-17936447
]
Branimir Lambov commented on CASSANDRA-20190:
---------------------------------------------
The first thing we need to start with is that we can't have a machine-dependent
on-disk order. Data shouldn't break if we move it from one machine type to
another. If I'm reading the above correctly, this is currently a problem and
needs to be corrected.
I personally don't have a strong preference which order to fix the data file to
(x86 and arm is where we are most used, so LE makes sense; on the other hand
conversion is so fast it's practically free, so switching to Java's default is
not really a loss). We should pick one and then we either introduce a flag to
flip it if someone runs the risk of a failure on upgrade, or, since this seems
to be a problem on the index summary which is not hard to recreate, maybe
recognize something's wrong on sstable load and throw away and recreate the
index file.
> MemoryUtil.setInt/getInt and similar use the wrong endianness
> -------------------------------------------------------------
>
> Key: CASSANDRA-20190
> URL: https://issues.apache.org/jira/browse/CASSANDRA-20190
> Project: Apache Cassandra
> Issue Type: Bug
> Components: Local/Other
> Reporter: Branimir Lambov
> Assignee: Dmitry Konstantinov
> Priority: Normal
> Time Spent: 1h
> Remaining Estimate: 0h
>
> `NativeCell`, `NativeClustering` and `NativeDecoratedKey` use the above
> methods from `MemoryUtil` to write and read data from native memory. As far
> as I can see they are meant to write data in big endian. They do not (they
> always correct to little endian).
> Moreover, they disagree with their `ByByte` versions on big-endian machines
> (which is only likely an issue on aligned-access architectures (x86 and arm
> should be fine)).
> The same is true for the methods in `Memory`, used by compression metadata as
> well as index summaries.
> We need to verify that this does not cause any problems, and to change the
> methods to behave as expected and document the behaviour by explicitly using
> `ByteOrder.LITTLE_ENDIAN` for any data that may have been persisted on disk
> with the wrong endianness.
>
> The current MemoryUtil behaviour (before the fix):
> ||Native
> order||MemoryUtil.setX||MemoryUtil.setXByByte||MemoryUtil.getX||MemoryUtil.getXByByte||
> |BE|LE|BE|LE|BE|
> |LE|LE|LE|LE|LE|
> shortly: MemoryUtil.setX/getX is LE, MemoryUtil.setXByByte/getXByByte is
> Native
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
