Actually the new TieredMergePolicy (only on trunk currently but I plan to backport for 3.2) lets you set the max merged segment size (maxMergedSegmentMB).
It's only an "estimate", but if it's set, it tries to pick a merge reaching around that target size. Mike http://blog.mikemccandless.com On Mon, May 2, 2011 at 9:03 AM, Shai Erera <ser...@gmail.com> wrote: > Hi > > Today, LogMP allows you to set different thresholds for segments sizes, > thereby allowing you to control the largest segment that will be > considered for merge + the largest segment your index will hold (=~ > threshold * mergeFactor). > > So, if you want to end up w/ say 20GB segments, you can set > maxMergeMB(ForOptimize) to 2GB and mergeFactor=10. > > However, this often does not achieve your desired goal -- if the index > contains 5 and 7 GB segments, they will never be merged b/c they are > bigger than the threshold. I am willing to spend the CPU and IO resources > to end up w/ 20 GB segments, whether I'm merging 10 segments together or > only 2. After I reach a 20GB segment, it can rest peacefully, at least > until I increase the threshold. > > So I wonder, first, if this threshold (i.e., largest segment size you > would like to end up with) is more natural to set than thee current > thresholds, > from the application level? I.e., wouldn't it be a simpler threshold to set > instead of doing weird calculus that depend on maxMergeMB(ForOptimize) > and mergeFactor? > > Second, should this be an addition to LogMP, or a different > type of MP. One that adheres to only those two factors (perhaps the > segSize threshold should be allowed to set differently for optimize and > regular merges). It can pick segments for merge such that it maximizes > the result segment size (i.e., don't necessarily merge in sequential > order), but not more than mergeFactor. > > I guess, if we think that maxResultSegmentSizeMB is more intuitive than > the current thresholds, application-wise, then this change should go > into LogMP. Otherwise, it feels like a different MP is needed, because > LogMP is already complicated and another threshold would confuse things. > > What do you think of this? Am I trying to optimize too much? :) > > Shai > > --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
