3 Apr 2012 18:46
Re: RFC: bloom filter (and generalized summary) support in leveldb
Dave Smith <dizzyd <at> dizzyd.com>
2012-04-03 16:46:39 GMT
2012-04-03 16:46:39 GMT
Hi Sanjay, Any update on the status of the bloom filter/summarization support? Any branches you'd like tested/exercised? Thanks, D. On Wed, Feb 29, 2012 at 4:20 PM, Sanjay Ghemawat <sanjay <at> google.com> wrote: >> If I understand the proposal aright, I believe that the Summarizer >> will only be deserialized when the index block (or metaindex block) is >> deserialized for the whole SST. Once that block is cached, you >> shouldn't have to worry about that cost again, unless you have >> insufficient room in your cache. > > Right, except that there should be no deserialization step. The summary > should be preformatted so it can be accessed quickly. > >> This does raise a random question -- would it be good to have a >> independent cache for index blocks? > > This can be achieved by setting "max_open_files". These many sstable files > are kept open and their index blocks (and associated summaries and bloom > filters in the future) will be kept cached in memory. So if you are getting > too many disk seeks to fetch index blocks, you can fix that by adjusting > max_open_files. It might be worthwhile to have independent controls for > max-open-files and index block caching, but I think it is not very urgent or > important to do that right now.(Continue reading)
RSS Feed