[
https://issues.apache.org/jira/browse/CASSANDRA-2635?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13042277#comment-13042277
]
Jonathan Ellis commented on CASSANDRA-2635:
-------------------------------------------
does setting dontneed on a read (our getDirectScanner) actually cause
pre-existing pages to be evicted post-read?
in other words, i'm wondering if this should really only change write behavior
and leave reads (for compaction) alone. (non-compaction reads are of course
not affected either way.)
also: suggest page_cache_hinting as the option name.
> make cache skipping optional
> ----------------------------
>
> Key: CASSANDRA-2635
> URL: https://issues.apache.org/jira/browse/CASSANDRA-2635
> Project: Cassandra
> Issue Type: Improvement
> Reporter: Peter Schuller
> Assignee: Peter Schuller
> Priority: Minor
> Attachments: CASSANDRA-2635-075.txt
>
>
> We've applied this patch locally in order to turn of page skipping; not
> completely but only for compaction/repair situations where it can be directly
> detrimental in the sense of causing data to become cold even though your
> entire data set fits in memory.
> It's better than completely disabling DONTNEED because the cache skipping
> does make sense and has no relevant (that I can see) detrimental effects in
> some cases, like when dumping caches.
> The patch is against 0.7.5 right now but if the change is desired I can make
> a patch for trunk. Also, the name of the configuration option is dubious
> since saying 'false' does not actually turn it off completely. I wasn't able
> to figure out a good name that conveyed the functionality in a short brief
> name however.
> A related concern as discussed in CASSANDRA-1902 is that the cache skipping
> isn't fsync:ing and so won't work reliably on writes. If the feature is to be
> retained that's something to fix in a different ticket.
> A question is also whether to retain the default to true or change it to
> false. I'm kinda leaning to false since it's detrimental in the "easy" cases
> of little data. In "big" cases with lots of data people will have to think
> and tweak anyway, so better to put the burden on that end.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
