Index time boosts, payloads, and long query strings
Hi , I'm relatively new to Solr/Lucene, and am using Solr (and not lucene directly) primarily because I can use it without writing java code (rest of my project is python coded). My application has the following requirements: (a) ability to search over multiple fields, each with different weight (b) If possible, I'd like to have the ability to add extra/diminished weights to particular tokens within a field (c) My query strings have large lengths (50-100 words) (d) My index is 500K+ documents 1) The way to (a) is field boosting (right?). My question is: Is all field boosting done at query time? Even if I give index time boosts to fields? Is there a performance advantage in boosting fields at index time vs at using something like fieldname:querystring^boost. 2) From what I've read, it seems that I can do (b) using payloads. However, as this link ( http://www.lucidimagination.com/blog/2009/08/05/getting-started-with-payloads/) suggests, I will have to write a payload aware Query Parser. Wanted to confirm if this is indeed the case - or is there a out-of-box way to implement payloads (am using Solr1.4) 3) For my project, the user fills multiple text boxes (for each query). I combine these into a single query (with different treatment for contents of each text box). Consequently, my query looks something like (fieldname1: queryterm1 queryterm2^2.0 queryterm3^3.0 +queryterm4)^1.0 Are there any guidelines for improving performance of such a system (sorry, this bit is vague) Any help with this will be great ! Girish Redekar http://girishredekar.net
Re: Index time boosts, payloads, and long query strings
Hi Erick - Maybe I mis-wrote. My question is: would "title:any_query^4.0" be faster/slower than applying index time boost to the field title. Basically, if I take *every* user query and search for it in title with boost (say, 4.0) - is it different than saying field title has boost 4.0? Cheers, Girish Redekar http://girishredekar.net On Sun, Nov 22, 2009 at 2:02 AM, Erick Erickson wrote: > I'll take a whack at index .vs. query boosting. They are expressing very > different concepts. Let's claim we're interested in boosting the title > field > > Index time boosting is expressing "this document's title is X more > important > > than a normal document title". It doesn't matter *what* the title is, > any query that matches on anything in this document's title will give this > document a boost. I might use this to give preferential treatment to all > encyclopedia entries or something. > > Query time boosting, like "title:solr^4.0" expresses "Any document with > solr > in > it's title is more important than documents without solr in the title". > This > really > only makes sense if you have other clauses that might cause a document > *without* > solr the title to match.. > > Since they are doing different things, efficiency isn't really relevant. > > HTH > Erick > > > On Sat, Nov 21, 2009 at 2:13 AM, Girish Redekar > wrote: > > > Hi , > > > > I'm relatively new to Solr/Lucene, and am using Solr (and not lucene > > directly) primarily because I can use it without writing java code (rest > of > > my project is python coded). > > > > My application has the following requirements: > > (a) ability to search over multiple fields, each with different weight > > (b) If possible, I'd like to have the ability to add extra/diminished > > weights to particular tokens within a field > > (c) My query strings have large lengths (50-100 words) > > (d) My index is 500K+ documents > > > > 1) The way to (a) is field boosting (right?). My question is: Is all > field > > boosting done at query time? Even if I give index time boosts to fields? > Is > > there a performance advantage in boosting fields at index time vs at > using > > something like fieldname:querystring^boost. > > 2) From what I've read, it seems that I can do (b) using payloads. > However, > > as this link ( > > > > > http://www.lucidimagination.com/blog/2009/08/05/getting-started-with-payloads/ > > ) > > suggests, I will have to write a payload aware Query Parser. Wanted to > > confirm if this is indeed the case - or is there a out-of-box way to > > implement payloads (am using Solr1.4) > > 3) For my project, the user fills multiple text boxes (for each query). I > > combine these into a single query (with different treatment for contents > of > > each text box). Consequently, my query looks something like (fieldname1: > > queryterm1 queryterm2^2.0 queryterm3^3.0 +queryterm4)^1.0 Are there any > > guidelines for improving performance of such a system (sorry, this bit is > > vague) > > > > Any help with this will be great ! > > > > Girish Redekar > > http://girishredekar.net > > >
Re: Index time boosts, payloads, and long query strings
Thanks Erick! After reading your answer, and re-reading the Solr wiki, I realized my folly. I used to think that index-time boosts when applied on a per-field basis are equivalent to query time boosts to that field. To ensure that my new understanding is correct , I'll state it in my words. Index time boosts will determine boost for a *document* if it is counted as a hit. Query time boosts give you control on boosting the occurrence of a query in a specific field. Please correct me if I'm wrong (again) :-) Girish Redekar http://girishredekar.net On Sun, Nov 22, 2009 at 8:25 PM, Erick Erickson wrote: > I still think they are apples and oranges. If you boost *all* titles, > you're effectively boosting none of them. Index time boosting > expresses "this document's title is more important than other > document titles." What I think you're after is "titles are more > important than other parts of the document. > > For this latter, you're talking query-time boosting. Boosting only > really makes sense if there are multiple clauses, something > like title:important OR body:unimportant. If this is true, speed > is irrelevant, you need correct behavior. > > Not that I think you'd notice either way. Modern computers > can do a LOT of FLOPS/sec. Here's an experiment: time > some queries (but beware of timing the very first ones, see > the Wiki) with boosts and without boosts. I doubt you'll see > enough difference to matter (but please do report back if you > do, it'll further my education ). > > But, depending on your index structure, you may get this > anyway. Generally, matches on shorter fields weigh more > in the score calculations than on longer fields. If you have > fields like title and body and you are querying on title:term OR > body:term, documents with term in the title will tend toward > higher scores. > > But before putting too much effort into this, do you have any > evidence that the default behavior is unsatisfactory? Because > unless and until you do, I think this is a distraction ... > > Best > Erick > > On Sun, Nov 22, 2009 at 8:37 AM, Girish Redekar > wrote: > > > Hi Erick - > > > > Maybe I mis-wrote. > > > > My question is: would "title:any_query^4.0" be faster/slower than > applying > > index time boost to the field title. Basically, if I take *every* user > > query > > and search for it in title with boost (say, 4.0) - is it different than > > saying field title has boost 4.0? > > > > Cheers, > > Girish Redekar > > http://girishredekar.net > > > > > > On Sun, Nov 22, 2009 at 2:02 AM, Erick Erickson > >wrote: > > > > > I'll take a whack at index .vs. query boosting. They are expressing > very > > > different concepts. Let's claim we're interested in boosting the title > > > field > > > > > > Index time boosting is expressing "this document's title is X more > > > important > > > > > > than a normal document title". It doesn't matter *what* the title is, > > > any query that matches on anything in this document's title will give > > this > > > document a boost. I might use this to give preferential treatment to > all > > > encyclopedia entries or something. > > > > > > Query time boosting, like "title:solr^4.0" expresses "Any document with > > > solr > > > in > > > it's title is more important than documents without solr in the title". > > > This > > > really > > > only makes sense if you have other clauses that might cause a document > > > *without* > > > solr the title to match.. > > > > > > Since they are doing different things, efficiency isn't really > relevant. > > > > > > HTH > > > Erick > > > > > > > > > On Sat, Nov 21, 2009 at 2:13 AM, Girish Redekar > > > wrote: > > > > > > > Hi , > > > > > > > > I'm relatively new to Solr/Lucene, and am using Solr (and not lucene > > > > directly) primarily because I can use it without writing java code > > (rest > > > of > > > > my project is python coded). > > > > > > > > My application has the following requirements: > > > > (a) ability to search over multiple fields, each with different > weight > > > > (b) If possible, I'd like to have the ability to add extra/diminished > > > > weights to particular tokens within a field > > > &g
Solr CPU usage
Hi I'm testing my Solr instance with multiple simultaneous requests. Here's my test. For an index of ~200K docs, I query Solr with 10 simultaneous threads. Can someone help me explain/improve the following observations: 1) Solr doesn't seem to use all the available CPU to improve response times (query times are good, but the time required to return documents aren't so good). My CPU seems to be running at ~30% 2) As expected, time for response increases as the num of requested results increase. What's surprising (and perplexing) is that Solr seems to use *more* of the CPU when I'm requesting *fewer* docs. Consequently, its performance in returning a larger result set is very bad 3) To counter 1, is there a way to make two Solr instances search on the same index (so that concurrent requests are served faster) Any help in this regard would be very useful. Thanks ! Girish Redekar http://girishredekar.net
Re: Solr CPU usage
Yonik, Am running both my server and client on ubuntu machines. The client is on a different box. The server CPU and RAM are both well below 50%. Girish Redekar http://girishredekar.net On Fri, Nov 27, 2009 at 10:07 PM, Yonik Seeley wrote: > On Fri, Nov 27, 2009 at 9:30 AM, Girish Redekar > wrote: > > Hi > > > > I'm testing my Solr instance with multiple simultaneous requests. Here's > my > > test. > > > > For an index of ~200K docs, I query Solr with 10 simultaneous threads. > Can > > someone help me explain/improve the following observations: > > > > 1) Solr doesn't seem to use all the available CPU to improve response > times > > (query times are good, but the time required to return documents aren't > so > > good). My CPU seems to be running at ~30% > > 2) As expected, time for response increases as the num of requested > results > > increase. What's surprising (and perplexing) is that Solr seems to use > > *more* of the CPU when I'm requesting *fewer* docs. Consequently, its > > performance in returning a larger result set is very bad > > This may point to another bottleneck - if the OS is low on free RAM, > it could be disk IO. > If this is on Windows, you could have contention reading the index files. > Otherwise, you may have a bottleneck in network IO. Is the client you > are testing with on the same box? > Is this Solr 1.4? > > -Yonik > http://www.lucidimagination.com >
Query time boosting with dismax
Hi, Is it possible to weigh specific query terms with a Dismax query parser? Is it possible to write queries of the sort ... field1:(term1)^2.0 + (term2^3.0) with dismax? Thanks, Girish Redekar http://girishredekar.net
