Andrew Wong has posted comments on this change. ( 
http://gerrit.cloudera.org:8080/15145 )

Change subject: WIP KUDU-1625: background op to GC ancient, empty rowsets
......................................................................


Patch Set 1:

> Patch Set 1:
>
> (4 comments)
>
> Did you explore building this heuristic into the compaction perf score and 
> letting the compaction op "delete" the rowsets? Curious what the trade-offs 
> are between this approach and that one.

Not seriously, because the merge compaction perf score heuristic is already 
quite complicated. We would need to fit deleted rowsets into the current perf 
score calculation, which assumes there are no deletes. We'd also likely need to 
make a decision on what level of write amplification we're willing to tolerate 
for the sake of deletes. My guess is that would be some user-specified flag 
(maybe per tablet/table?). I took a brief look over 
https://docs.google.com/document/u/1/d/1yTfxt0_2p5EfIjCnjJCt3o-nB9xk-Kl2O8yKTA1LQrQ/edit
 and left some comments on the ticket, but I think it'd require more rigorous 
thought.

We could consider doing this with MajorDeltaCompactions, but right now I think 
those only try to compact a single rowset at a time, which isn't great either.

Some obvious tradeoffs are that this approach _only_ affects entirely empty 
rowsets, which might not be too common. But it's significantly simpler.


--
To view, visit http://gerrit.cloudera.org:8080/15145
To unsubscribe, visit http://gerrit.cloudera.org:8080/settings

Gerrit-Project: kudu
Gerrit-Branch: master
Gerrit-MessageType: comment
Gerrit-Change-Id: I696e2a29ea52ad4e54801b495c322bc371787124
Gerrit-Change-Number: 15145
Gerrit-PatchSet: 1
Gerrit-Owner: Andrew Wong <[email protected]>
Gerrit-Reviewer: Adar Dembo <[email protected]>
Gerrit-Reviewer: Alexey Serbin <[email protected]>
Gerrit-Reviewer: Andrew Wong <[email protected]>
Gerrit-Reviewer: Kudu Jenkins (120)
Gerrit-Reviewer: Tidy Bot (241)
Gerrit-Comment-Date: Fri, 31 Jan 2020 20:33:26 +0000
Gerrit-HasComments: No

Reply via email to