[ 
https://issues.apache.org/jira/browse/OAK-5058?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15690347#comment-15690347
 ] 

Michael Dürig commented on OAK-5058:
------------------------------------

The more sever problem with 2. is that it will defer running gc on big 
repositories for too long. Technically I really think 1. is the right solution 
as we want to run gc once a certain amount of garbage has piled up. The 
absolute growth of the repository should be a good (and simple) estimator here. 

3. is more about user's expectations re. gc running regularly. Without 3. gc 
would probably be skipped a couple of times on a fresh instance. Not a problem 
(rather a feature IMO) but still I think users would complain...

> Improve GC estimation strategy based on both absolute size and relative 
> percentage
> ----------------------------------------------------------------------------------
>
>                 Key: OAK-5058
>                 URL: https://issues.apache.org/jira/browse/OAK-5058
>             Project: Jackrabbit Oak
>          Issue Type: Improvement
>          Components: segment-tar
>    Affects Versions: 1.5.12
>            Reporter: Andrei Dulceanu
>            Assignee: Andrei Dulceanu
>            Priority: Minor
>             Fix For: 1.6, 1.5.15
>
>         Attachments: OAK-5058-01.patch
>
>
> A better way of deciding whether GC should run or not might be by looking at 
> the numbers computed in {{SizeDeltaGcEstimation}} from both an absolute size 
> and relative percentage point of view. For example it would make sense to  
> run compaction only if at least one criterion is met: "run if there is > 50% 
> increase or more than 10GB".
> Since the absolute threshold is already implemented (see 
> {{SegmentGCOptions.SIZE_DELTA_ESTIMATION_DEFAULT}}), it would be nice to add 
> also something like {{SegmentGCOptions.SIZE_PERCENTAGE_ESTIMATION_DEFAULT}} 
> and use it in making the decision.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to