lucene-java-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Nikhil Chhaochharia <nikhil...@yahoo.com>
Subject Re: Multiple Indices vs Single Index
Date Fri, 21 Sep 2007 05:37:04 GMT
Thanks Grant and Chris for the replies.

I am looking at a single index because the 40 index system has started having performance
issues at high load.  My daily traffic is increasing at a steady pace and about 40% of the
traffic is concentrated in a 2 hour period and searches start slowing down just a little bit.

All my indices have exactly the same fields, so I guess fieldcache and sorting are not an
issue.  I do not do any updates, so I just open 40 searchers and store them in a HashMap.
 When a query comes, I select the appropriate searcher and fire the query.  So code maintainance
also is not an issue.

The different IDF statistics are a valid point.  I have not analysed the difference in the
results returned - I sort of assumed the results will be same in both cases.  I will look
at the difference in results.


There is one more point.  Sooner or later, I will have to move to multiple servers due to
increasing traffic.  (I will be handling 30,000+ hits per hour by the end of the year)  One
option is to have full index on all servers and load-balance them.  Another is to have half
the indices on one server and half of them on the other.  The front-end (separate server)
will then fire the query on the appropriate server.  Any suggestions on which one would be
a better choice ?  All data on all servers will give me redundancy, the system will be up
even if one server goes down.  Also, adding more servers would be trivial.


Thanks,
Nikhil


----- Original Message ----
From: Chris Hostetter <hossman_lucene@fucit.org>
To: java-user@lucene.apache.org
Sent: Friday, 21 September, 2007 4:02:38 AM
Subject: Re: Multiple Indices vs Single Index


: I was wondering if it will be better to just have 1 large index with all 
: the 40 indices combined.  I do not need to do dual-queries and my total 
: index size (if I create a single index) is about 3.4GB.  It will 
: increase to maximum of 5-6 GB.  I am running this on a dedicated machine 
: with 8GB RAM.

off the top of my head, there are 3 main reasons i can think of that would 
motivate one choice over another -- ultimately it's up to you...

1) FieldCache and sorting ... if all 40 sets of of Documents contain 
have consistently named fields, then there won't be much difference 
between 40 indexes and 1 index ... but if each of those 40 sets contain 
documents with radically differnet fields -- and you want to sort on N 
differnet fields for each sets -- then the total FieldCache sizes for each 
of those 40 indexes will be smaller then the FieldCaches for one gian 
index (because every document will get an entry wethe it makes sense or 
not.

2) idf statistics.  if you have common fields you search regardless of 
document set, the 40 index approach will maintain seperate sttistics -- 
this may be important if some terms are very common in only som docsets.  
the word "albino" may be really common in docset A but only one doc in 
docset B has it ... in the 40 index appraoch querying B for (albino 
elephant) will give a lot of weight to albino because it's so rare, but in 
the single index case albino may not be considered as significant because 
of ht unified idf value for all docsets 9even if hte query is constrained 
to docset B) ... again: this only matters if the fields overlap, if every 
docset has a unique set of fields then the idfs will be unique because 
they are by field)

3) management: it's probably a lot simpler to maintain and manage code 
that deals with one index then code that deals with 40 indexes. you milage 
may vary.


-Hoss


---------------------------------------------------------------------
To unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-user-help@lucene.apache.org





---------------------------------------------------------------------
To unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-user-help@lucene.apache.org


Mime
View raw message