[jira] [Commented] (CASSANDRA-9829) Dynamically adjust LCS level sizes
[ https://issues.apache.org/jira/browse/CASSANDRA-9829?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17715858#comment-17715858 ] Jonathan Ellis commented on CASSANDRA-9829: --- Do we think UCS replaces substantially all LCS use cases, or just many of them? > Dynamically adjust LCS level sizes > -- > > Key: CASSANDRA-9829 > URL: https://issues.apache.org/jira/browse/CASSANDRA-9829 > Project: Cassandra > Issue Type: New Feature > Components: Local/Compaction >Reporter: Jonathan Ellis >Priority: Normal > Labels: compaction, lcs, performance > Fix For: 5.x > > > LCS works best when the top level is full. Then 90% of reads can be served > from a single sstable. By contrast if the top level is only 10% full then > 90% of reads will be served from two. This results in worse performance as > well as confused users. > To address this, we can adjust the ideal top level size to how much data is > actually in it (and set each corresponding lower level to 1/10 of the next > one above). > (This is an idea [from > rocksdb|https://www.reddit.com/r/IAmA/comments/3de3cv/we_are_rocksdb_engineering_team_ask_us_anything/ct4asen].) -- 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
[jira] [Commented] (CASSANDRA-9829) Dynamically adjust LCS level sizes
[ https://issues.apache.org/jira/browse/CASSANDRA-9829?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17715708#comment-17715708 ] Michael Semb Wever commented on CASSANDRA-9829: --- [~jbellis], [~blambov], do we still wish to keep this ticket open with CEP-26 (UCS: CASSANDRA-18397) landing in 5.0 ? > Dynamically adjust LCS level sizes > -- > > Key: CASSANDRA-9829 > URL: https://issues.apache.org/jira/browse/CASSANDRA-9829 > Project: Cassandra > Issue Type: New Feature > Components: Local/Compaction >Reporter: Jonathan Ellis >Priority: Normal > Labels: compaction, lcs, performance > Fix For: 5.x > > > LCS works best when the top level is full. Then 90% of reads can be served > from a single sstable. By contrast if the top level is only 10% full then > 90% of reads will be served from two. This results in worse performance as > well as confused users. > To address this, we can adjust the ideal top level size to how much data is > actually in it (and set each corresponding lower level to 1/10 of the next > one above). > (This is an idea [from > rocksdb|https://www.reddit.com/r/IAmA/comments/3de3cv/we_are_rocksdb_engineering_team_ask_us_anything/ct4asen].) -- 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
