Regexes where not optimized against backtracking as I did not know. A typical one looked like this :
-.*\.domain\.*.* I guess something like this would be better and give less backtracking : http:\/\/www\.domain\..* As for the automaton no reason a priori since I really never looked at it. Does it used TRIE like pattern matching ? Which would be very fast and appropriate I guess. I will have a go at it and see how it helps. Thx. 2011/6/4 Julien Nioche <[email protected]> > As Kirby pointed out the automaton-based filter should be far more efficent > + its syntax is more restricted than the regex one but not dissimilar. > What do your filters look like? Any reason why you can't use the automaton > instead? > > Julien > > On 4 June 2011 09:44, MilleBii <[email protected]> wrote: > > > Just for the record the impact can be very very bad if you add too many > > regexes. I just finished a test and I got a factor 20 slower just for the > > generate step by adding 30 or so regexes in the filter. So beware. > > > > 2011/6/3 MilleBii <[email protected]> > > > > > Indeed I'm running a vertical search engine too however I want to > improve > > > it as well on several front. > > > Scoring is one way but it does not prevent uninteresting content to > > > creep-in the crawld which eventually grows big/too big wasting > ressources > > > for nothing. I > > > > > > Filtering is another way at the cost of a lot of regexes, hence this > > > question. > > > > > > Third I see crawldb pruning you want to ditch all urls that are below > a > > > certain score. A question I asked a long time ago and the answer was > > write > > > your own mapred for that, a bit too far fetch for me insofar. > > > > > > What would be top for me is to be able to extract properties about > > > pages/urls in whatever phase scoring, indexing and be able to use those > > > properties during the generate phase as a kind of feedback loop. It is > a > > > real pain to be forced to try to merge this information into a score > > > > > > > > > > > > 2011/6/2 Kirby Bohling <[email protected]> > > > > > >> I see from your e-mails that you are modifying the Scoring algorithm, > > >> the only other option I see is to write the scoring algorithm which > > >> detects that this is content you don't want to crawl, and this lowers > > >> the score. As I recall, links with the highest score are crawled > > >> first, so in the end that might be easier. Which sounds like it'd be > > >> writing a Vertical Search engine of some type (either that, or Spam > > >> detector with your personal/custom definition of Spam). > > >> > > >> I know several people on the this list or the dev list are writing > > >> vertical search engines, maybe they would have more thoughts or info. > > >> > > >> Kirby > > >> > > >> On Thu, Jun 2, 2011 at 3:47 PM, MilleBii <[email protected]> wrote: > > >> > Yes I remember reading that a few years ago. > > >> > But frankly I can't design by hand such finite automaton which will > > ever > > >> > changing by the way. > > >> > > > >> > Even adding regexes by hand, is most likely a daunting task for me. > > >> > > > >> > 2011/6/2 Kirby Bohling <[email protected]> > > >> > > > >> >> From what I remember of earlier advise, you really want to use the > > >> >> Automaton filter if at all possible, rather than series of straight > > >> >> regex. Using the Automaton should be linear with respect to the > > >> >> number of characters in the URL. Building the actual automaton > could > > >> >> be fairly time consuming, but as you'll re-using it often, likely > > >> >> worth the cost. > > >> >> > > >> >> > > >> >> > > >> > > > http://nutch.apache.org/apidocs-1.2/org/apache/nutch/urlfilter/automaton/package-summary.html > > >> >> > > >> >> A series of Java Regex's should also be linear with the number of > > >> >> characters in the URL assuming you avoid specific constructs (the > > >> >> things that cause back tracking, where it effectively tries to > ensure > > >> >> that one group/subgroup is equal to a later group/subgroup is the > > >> >> primary culprit). Each Regex will add to the constant multiple in > > >> >> front of the number of characters. > > >> >> > > >> >> I've used the Automaton library, and if you can work within the > > >> >> limitations (it is a classic regex matcher with limited operators > > >> >> relative to say Perl 5 Compatible Regular Expressions). > > >> >> > > >> >> I don't have any practical experience with Nutch for a large scale > > >> >> crawl, but based upon my experience with using regular expressions > > and > > >> >> the Automaton Library, I know it is much faster. I recall Andrej > > >> >> talking about it being much faster. It might also be worth while > for > > >> >> Nutch to look into Lucene's optimized versions of Automaton (they > > >> >> ported over several critical operations for use in Lucene and the > > >> >> Fuzzy matching when computing the Levenshtien distance). > > >> >> > > >> >> I can't seem to find the thread where I saw that advice given, but > > you > > >> >> can see the thread where they discuss adding the Automaton URL > filter > > >> >> back in Nutch 0.8 and it seems to agree with my experience in using > > >> >> both. > > >> >> > > >> >> > > >> >> > > >> > > > http://lucene.472066.n3.nabble.com/Much-faster-RegExp-lib-needed-in-nutch-td623308.html > > >> >> > > >> >> Kirby > > >> >> > > >> >> > > >> >> > > >> >> On Thu, Jun 2, 2011 at 2:42 PM, MilleBii <[email protected]> > wrote: > > >> >> > What will be the impact of a growing big regex-urlfilter ? > > >> >> > > > >> >> > I ask this because there are more & more sites that I want to > > filter > > >> >> out, > > >> >> > it will limit the # of unecessary pages at a cost of lots of url > > >> >> > verification. > > >> >> > Side question since I already have pages from those sites in the > > >> crawldb, > > >> >> > will they be removed ever ? What would be the method to remove > them > > ? > > >> >> > > > >> >> > -- > > >> >> > -MilleBii- > > >> >> > > > >> >> > > >> > > > >> > > > >> > > > >> > -- > > >> > -MilleBii- > > >> > > > >> > > > > > > > > > > > > -- > > > -MilleBii- > > > > > > > > > > > -- > > -MilleBii- > > > > > > -- > * > *Open Source Solutions for Text Engineering > > http://digitalpebble.blogspot.com/ > http://www.digitalpebble.com > -- -MilleBii-
