Thanks for the explanation regarding IntersectTermsEnum.next(). I understand better now.
How about indexing a string field containing the firstname and lastname of a person or possibly its address. Values for these fields have variable length and usually they have more than 16 bytes. If my understanding is correct then they should (can) not be indexed as points. Range queries on this fields appear mostly when users implement custom pagination (without using searchAfter) on the result set. The alternative that I consider and seems to work well for the use-cases I mentioned is using doc values and SortedDocValuesField.newSlowRangeQuery. Στις Τρί, 2 Απρ 2019 στις 3:01 μ.μ., ο/η Robert Muir <rcm...@gmail.com> έγραψε: > Can you explain a little more about your use-case? I think that's the > biggest problem here for term range query. Pretty much all range > use-cases are converted over to space partitioned data structures > (Points) so its unclear why this query would even be used for anything > serious. > > To answer your question, IntersectTermsEnum.next() doesn't iterate > over the whole index, just the term dictionary. But the lines are > blurred since if there's only one posting for a term, that single > posting is inlined directly into the term dictionary.... and often a > ton of terms fit into this category, so seeing 80% time in the term > dictionary wouldn't surprise me. > > On Tue, Apr 2, 2019 at 8:03 AM Stamatis Zampetakis <zabe...@gmail.com> > wrote: > > > > Thanks Robert and Uwe for your feedback! > > > > I am not sure what you mean that the slowness does not come from the > iteration over the term index. > > I did a small profiling (screenshot attached) of sending repeatedly a > TermRangeQuery that matches almost the whole index and I observe that 80% > of the time is spend on IntersectTermsEnum.next() method. > > Isn't this the method which iterates over the index? > > > > > > > > Στις Δευ, 1 Απρ 2019 στις 6:57 μ.μ., ο/η Uwe Schindler <u...@thetaphi.de> > έγραψε: > >> > >> Hi again, > >> > >> > The problem with TermRangeQueries is actually not the iteration over > the > >> > term index. The slowness comes from the fact that all terms between > start > >> > and end have to be iterated and their postings be fetched and those > postings > >> > be merged together. If the "source of terms" for doing this is just a > simple > >> > linear iteration of all terms from/to or the automaton intersection > does not > >> > really matter for the query execution. The change to prefer the > automaton > >> > instead of a simple term iteration is just to allow further > optimizations, for > >> > more info see https://issues.apache.org/jira/browse/LUCENE-5879 > >> > >> So the above issue brings no further improvements (currently). So as > said before, a simple enumeration of all all terms and fetching their > postings will be as fast. By using an automaton, the idea here is that the > term dictionary may have some "percalculated" postings list for prefix > terms. So instead of iterating over all terms and merging their postings, > the terms dictionary could return a "virtual term" that contains all > documents for a whole prefix and store the merged postings list in the > index file. This was not yet implemented but is the place to hook into. You > could create an improved BlockTermsDict implementation, that allows to get > a PostingsEnum for a whole prefix of terms. > >> > >> Not sure how much of that was already implemented by LUCENE-5879, but > it allows to do this. So here is where you could step in and improve the > terms dictionary! > >> > >> Uwe > >> > >> > Uwe > >> > > >> > ----- > >> > Uwe Schindler > >> > Achterdiek 19, D-28357 Bremen > >> > http://www.thetaphi.de > >> > eMail: u...@thetaphi.de > >> > > >> > > -----Original Message----- > >> > > From: Robert Muir <rcm...@gmail.com> > >> > > Sent: Monday, April 1, 2019 6:30 PM > >> > > To: java-user <java-user@lucene.apache.org> > >> > > Subject: Re: Clarification regarding BlockTree implementation of > >> > > IntersectTermsEnum > >> > > > >> > > The regular TermsEnum is really designed for walking terms in > linear order. > >> > > it does have some ability to seek/leapfrog. But this means paths in > a query > >> > > automaton that match no terms result in a wasted seek and cpu, > because > >> > the > >> > > api is designed to return the next term after regardless. > >> > > > >> > > On the other hand the intersect() is for intersecting two automata: > query > >> > > and index. Presumably it can also remove more inefficiencies than > just the > >> > > wasted seeks for complex wildcards and fuzzies and stuff, since it > can > >> > > "see" the whole input as an automaton. so for example it might be > able to > >> > > work on blocks of terms at a time and so on. > >> > > > >> > > On Mon, Apr 1, 2019, 12:17 PM Stamatis Zampetakis > >> > <zabe...@gmail.com> > >> > > wrote: > >> > > > >> > > > Yes it is used. > >> > > > > >> > > > I think there are simpler and possibly more efficient ways to > implement a > >> > > > TermRangeQuery and that is why I am looking into this. > >> > > > But I am also curious to understand what IntersectTermsEnum is > >> > supposed > >> > > to > >> > > > do. > >> > > > > >> > > > Στις Δευ, 1 Απρ 2019 στις 5:34 μ.μ., ο/η Robert Muir < > rcm...@gmail.com> > >> > > > έγραψε: > >> > > > > >> > > > > Is this IntersectTermsEnum really being used for term range > query? > >> > > Seems > >> > > > > like using a standard TermsEnum, seeking to the start of the > range, then > >> > > > > calling next until the end would be easier. > >> > > > > > >> > > > > On Mon, Apr 1, 2019, 10:05 AM Stamatis Zampetakis > >> > > <zabe...@gmail.com> > >> > > > > wrote: > >> > > > > > >> > > > > > Hi all, > >> > > > > > > >> > > > > > I am currently working on improving the performance of range > >> > queries > >> > > on > >> > > > > > strings. I've noticed that using TermRangeQuery with > low-selective > >> > > > > queries > >> > > > > > is a very bad idea in terms of performance but I cannot > clearly explain > >> > > > > why > >> > > > > > since it seems related with how the IntersectTermsEnum#next > >> > method > >> > > is > >> > > > > > implemented. > >> > > > > > > >> > > > > > The Javadoc of the class says that the terms index (the > burst-trie > >> > > > > > datastructure) is not used by this implementation of > TermsEnum. > >> > > > However, > >> > > > > > when I see the implementation of the next method I get the > >> > impression > >> > > > > that > >> > > > > > this is not accurate. Aren't we using the trie structure to > skip parts > >> > > > of > >> > > > > > the data when the automaton states do not match? > >> > > > > > > >> > > > > > Can somebody provide a high-level intutition of what > >> > > > > > IntersectTermsEnum#next does? Initially, I thought that it is > >> > > > traversing > >> > > > > > the whole trie structure (skipping some branches when > necessary) but > >> > I > >> > > > > may > >> > > > > > be wrong. > >> > > > > > > >> > > > > > Thanks in advance, > >> > > > > > Stamatis > >> > > > > > > >> > > > > > >> > > > > >> > >> > >> --------------------------------------------------------------------- > >> To unsubscribe, e-mail: java-user-unsubscr...@lucene.apache.org > >> For additional commands, e-mail: java-user-h...@lucene.apache.org > >> > > > > --------------------------------------------------------------------- > > To unsubscribe, e-mail: java-user-unsubscr...@lucene.apache.org > > For additional commands, e-mail: java-user-h...@lucene.apache.org > > --------------------------------------------------------------------- > To unsubscribe, e-mail: java-user-unsubscr...@lucene.apache.org > For additional commands, e-mail: java-user-h...@lucene.apache.org > >
