Would our approach to limit the search top 250 documents (and then sort these 250 documents) work fine ? Or even 250 unique terms with a lot of users is bad on memory when sorting ?
We didn't look at trie fields - I will do though, thanks for the tip ! We do store the original 'Data' field (only the 'SearchableData' field is analyzed, all other fields are not analyzed), the users mainly sort on numeric values; not a lot on string values (in fact I could compltely drop the sort by string feature). We do pad our numbers with zeros though (for example: 10 becomes 00000010, etc.) because we had trouble with sorting (100 was smaller than 2) ; is that considered as "string sorting" ? This might explain a part of the problem. Why/how would I reduce the count of unique terms? - Mike aka...@gmail.com On Tue, Aug 17, 2010 at 3:28 PM, Erick Erickson <erickerick...@gmail.com>wrote: > If you have tens of millions of documents, almost all with unique fields > that you're sorting on, you'll chew through memory like there's no > tomorrow. > > Have you looked at trie fields? See: > > http://www.lucidimagination.com/blog/2009/05/13/exploring-lucene-and-solrs-trierange-capabilities/ > > I'm a little concerned that the user can sort on Data. Any field used for > sorting > should NOT be analyzed, so unless you are indexing "Data" unanalyzed, > that's > a problem. And if you are sorting on strings unique to each document, > that's > also a memory hog. Not to mention whether capitalization counts. > > You might enumerate the terms in your index for each of the sortable fields > to figure out what the total number of unique terms each is and use that as > a basis for reducing their count.... > > HTH > Erick > > On Tue, Aug 17, 2010 at 3:05 PM, Michel Nadeau <aka...@gmail.com> wrote: > > > Hi Erick, > > > > Here's some more details about our structure. First here's an example of > > document in our index : > > > > PrimaryKey = SJAsfsf353JHGada66GH6 (it's a hash) > > DocType = X > > Data = This is the data > > SearchableContent = This is the data > > DateCreated = <timestamp> > > DateModified = <timestamp> > > Counter1 = 17 > > Counter2 = 3 > > Average = 0.17 > > Cost = 200 > > > > The users are able to sort on almost all fields: Data, DateCreated, > > DateModified, Counter1, Counter2, Average, Cost. > > > > When we search, we always search on the 'SearchableContent' field and we > > have at least one filter on the DocType (because we have many document > > types > > in the same index). So a common search that would find the document above > > is > > "data *AND DocType:X*" (we automatically add the "*AND DocType:X*" part > > using Lucene Filters. > > > > I would say that the number of unique terms in the field being sorted on > is > > very big - for example timestamps, almost all unique, counters, average, > > cost, data... so if a query finds 10M results, it's almost 10M different > > values to sort. About cache and warm-up queries : we don't use warm-up > > queries -at all- because we have absolutely no idea of what users are > going > > to search for (they can search for absolutely anything). About "returning > > 10M" documents, right, we don't actually return the 10M documents, we use > > pagination to return documents X to Y of the 10M (and the 10M was only an > > example, it can be anywhere between 1K and 100M results). The pagination > > usually works fine and fast, our problem is really sorting. > > > > Our "Lucene Reader" process has 2GB of ram allowed, here's how I start it > - > > > > java -Xmx2048m -jar LuceneReader.jar > > > > The problem really seems to be a ram problem, but I can't be 100% sure > (any > > help about how to be sure is welcome). > > > > Our current idea of a solution would be to get maximum 250 results (the > > more > > relevant ones; more results than that is totally useless in our system) > so > > the sort should work fine on a small data set like that, but we want to > > make > > sure we're doing everything right before doing that so we don't run in > the > > same problems again. > > > > Thank you very much; let me know if you need any more details! > > > > - Mike > > aka...@gmail.com > > > > > > On Mon, Aug 16, 2010 at 4:01 PM, Erick Erickson <erickerick...@gmail.com > > >wrote: > > > > > Let's back up a minute. The number of matched records is not > > > important when sorting, what's important is the number of unique > > > terms in the field being sorted. Do you have any figures on that? One > > > very common sorting issue is sorting on very fine date time > resolutions, > > > although your examples don't include that... > > > > > > Now, cache loading is an issue. The very first time you sort on a > field, > > > all the values are read into a cache. Is it possible this is the source > > > of your problems? You can cure this with warmup queries. The take-away > > > is that measuring the response time for the first sorted query is > > > very misleading. > > > > > > Although if you're sorting on many, many, many email addresses, > > > that could be "interesting". > > > > > > The comment "returning 10,000,000 documents" is, I hope, a > > > misstatement. If you're trying to *return* 10M docs sorting > > > is irrelevant compared to assembling that many documents. If > > > you're trying to return the first 100 of 10M documents, it should > > > work. > > > > > > Overall, we need more details on what you're sorting and what > > > you're trying to return as well as how you're measuring before > > > we can say much.... > > > > > > Along with how much memory you're giving your JVM to work with, > > > what "exploding" means. Are you CPU bound? IO bound? Swapping? > > > You need to characterize what is going wrong before worrying about > > > solutions...... > > > > > > Best > > > Erick > > > > > > On Mon, Aug 16, 2010 at 3:08 PM, Michel Nadeau <aka...@gmail.com> > wrote: > > > > > > > Hi, > > > > > > > > we are building an application using Lucene and we have HUGE data > sets > > > (our > > > > index contains millions and millions and millions of documents), > which > > > > obviously cause us very important problems when sorting. In fact, we > > > > disabled sorting completely because the servers were just exploding > > when > > > > trying to sort results in RAM. The users using the system can search > > for > > > > whatever they want, so we never know how many results will be > returned > > - > > > a > > > > search can return 10 documents (no problem with sorting) or > 10,000,000 > > > > (huge > > > > sorting problems). > > > > > > > > I woke up this morning and had a flash : is it possible with Lucene > to > > > have > > > > a "natural sorting" of documents? For example, let's say I have 3 > > columns > > > I > > > > want to be able to sort by : first name, last name, email; I would > have > > 3 > > > > different indexes with the very same data but with a different > primary > > > key > > > > for sorting. I know it's far fetched, and I have never seen anything > > like > > > > that since I use Lucene, but we're just desperate... how people do to > > > have > > > > huge data sets, a lot of users, and sort!? > > > > > > > > Thanks, > > > > > > > > Mike > > > > > > > > > >
