[ https://issues.apache.org/jira/browse/CASSANDRA-11106?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15130282#comment-15130282 ]
Marcus Eriksson commented on CASSANDRA-11106: --------------------------------------------- [~kohlisankalp] true, converted it to a subtask of this since we are touching those parts here > Experiment with strategies for picking compaction candidates in LCS > ------------------------------------------------------------------- > > Key: CASSANDRA-11106 > URL: https://issues.apache.org/jira/browse/CASSANDRA-11106 > Project: Cassandra > Issue Type: Improvement > Reporter: Marcus Eriksson > Labels: lcs > > Ideas taken here: http://rocksdb.org/blog/2921/compaction_pri/ > Current strategy in LCS is that we keep track of the token that was last > compacted and then we start a compaction with the sstable containing the next > token (kOldestSmallestSeqFirst in the blog post above) > The rocksdb blog post above introduces a few ideas how this could be improved: > * pick the 'coldest' sstable (sstable with the oldest max timestamp) - we > want to keep the hot data (recently updated) in the lower levels to avoid > write amplification > * pick the sstable with the highest tombstone ratio, we want to get > tombstones to the top level as quickly as possible. -- This message was sent by Atlassian JIRA (v6.3.4#6332)