Le lun. 24 avr. 2017 à 05:02, 马可阳 <makey...@jd.com> a écrit :
> Last week I did a test: > Use a pk to search elasticsearch, and it’s TPS is 16000+(I use mmap fs, > filter etc to optimize search performance) while I use Redis to do it and > it’s TPS is 80000+ > If you want optimal performance, you should use the GET API. The search API introduces overhead like opening index inputs, creating weights, interacting with the query cache, etc. which you can skip if you use the GET API. It can also apply deleted docs more efficiently since it knows there is a at most one match for any key. There are a bouch of overhead like network and other staff I’ll exclude > later, but I wonder if Lucene can build a hash index to support this > scenario? > I believe you could implement a custom PostingsFormat that looks up terms using hashing. It would have some limitations, eg. it would not be possible to run prefix or range queries, but for a primary key this should not be a problem. It could be an interesting addition to the Lucene/codecs module. However I need to warn you that it is quite unlikely that we end up supporting another postings format from a backward compatibility perspective as it is a lot of efforts so we only want to support one.
