[ https://issues.apache.org/jira/browse/CASSANDRA-11119?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15131501#comment-15131501 ]
Jonathan Ellis edited comment on CASSANDRA-11119 at 2/11/16 5:01 PM: --------------------------------------------------------------------- We can actually get pretty accurate with STCS thanks to the bucketing we do anyway to compute next-task-to-compact. For LCS we're probably still looking at some pretty rough estimates. was (Author: jbellis): We can actually get pretty accurate with STCS thanks to HyperLogLog. For LCS we're probably still looking at some pretty rough estimates. > Add bytes-to-compact estimate > ----------------------------- > > Key: CASSANDRA-11119 > URL: https://issues.apache.org/jira/browse/CASSANDRA-11119 > Project: Cassandra > Issue Type: Improvement > Components: Compaction > Reporter: Jonathan Ellis > Fix For: 3.x > > > Pending tasks is essentially useless at letting operators know how far behind > compaction is. For STCS, because compaction tasks can vary in size by > several orders of magnitude, so "4 more tasks" is a very wide range. For > LCS, I think the problem is that we only look at first-order work (how much > larger is level X than it "should" be?) and not cascading effects, > particularly in L0 (if we push this into L1, how much does that push into L2 > etc?). > I propose adding an estimate of bytes-to-compact (and deprecating estimated > tasks). -- This message was sent by Atlassian JIRA (v6.3.4#6332)