[
https://issues.apache.org/jira/browse/CASSANDRA-9830?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15089775#comment-15089775
]
Paulo Motta commented on CASSANDRA-9830:
----------------------------------------
bq. The cstar runs that you kicked off didn't work because you aren't on the
list of repos for cstar, so I kicked off a new ssd test. The increase looks
modest, but there is an improvement.
Thanks for triggering the test again!
bq. Also, is there any way to get memory usage during the tests?
The last {{nodetool tablestats}} command should print memory usage stats in the
console, but for some reason the console output is unavailable. Is there any
way to retrieve it manually [~enigmacurry]?
bq. I think we should be skipping creating the bloom filter for the leveled
major compaction as well. That's because in major compaction, while we aren't
always adding at what ends up being the highest level after we are done, we are
always writing the highest level for a given key. Plus, this will ensure that
whichever level ends up as the highest will not have bloom filters.
Thanks for the feedback. That's right, we should definitely support this in
major leveled compaction. I will update the patch post back soon.
> Option to disable bloom filter in highest level of LCS sstables
> ---------------------------------------------------------------
>
> Key: CASSANDRA-9830
> URL: https://issues.apache.org/jira/browse/CASSANDRA-9830
> Project: Cassandra
> Issue Type: New Feature
> Components: Compaction
> Reporter: Jonathan Ellis
> Assignee: Paulo Motta
> Priority: Minor
> Labels: performance
> Fix For: 3.x
>
>
> We expect about 90% of data to be in the highest level of LCS in a fully
> populated series. (See also CASSANDRA-9829.)
> Thus if the user is primarily asking for data (partitions) that has actually
> been inserted, the bloom filter on the highest level only helps reject
> sstables about 10% of the time.
> We should add an option that suppresses bloom filter creation on top-level
> sstables. This will dramatically reduce memory usage for LCS and may even
> improve performance as we no longer check a low-value filter.
> (This is also an idea from RocksDB.)
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
