import pymongo import pymongo.errors import trafilatura import trafilatura.feeds import trafilatura.sitemaps import trafilatura.spider import trafilatura.utils import trafilatura.external import sys import courlan import urllib from datetime import datetime import click import logging as LOGGER import os import pprint import re import time import collections import math import random import hashlib LANGUAGE= os.getenv("SUCKER_LANGUAGE","sk") DOMAIN = os.getenv("SUCKER_DOMAIN","sk") BATCHSIZE=int(os.getenv("SUCKER_BATCHSIZE","10")) CONNECTION=os.getenv("SUCKER_CONNECTION","mongodb://root:example@localhost:27017/") DBNAME=os.getenv("SUCKER_DBNAME","crawler") MINFILESIZE=300 MAXFILESIZE=10000000 MINTEXTSIZE=200 CHECK_PARAGRAPH_SIZE=150 TEXT_TRASH_SIZE=200 TEXT_TRASH_RATIO=0.6 def split_train(res): trainset = [] testset = [] for i,item in enumerate(res): if i % 10 == 0: testset.append(item) else: trainset.append(item) return trainset,testset def calculate_checksums(text): """ @return fingerprints of a paragraphs in text. Paragraphs are separated by a blank line """ checksums = [] sizes = [] hval = 0 hsz = 0 sz = 0 for c in text: cv = ord(c) sz += 1 if cv > 64: # ignore non-ascii hval += (hval << 3) + cv zv = hval >> 31 hval &= 0x7fffffff hval += zv hsz += 1 if c == "\n" and hsz > 0: if hsz > CHECK_PARAGRAPH_SIZE: checksums.append(hval) sizes.append(sz) sz = 0 hsz = 0 if hsz > CHECK_PARAGRAPH_SIZE: checksums.append(hval) sizes.append(sz) return checksums, sizes def is_robot_good(link,rules): # check robots.txt rules if rules is not None and not rules.can_fetch("*", link): return False return True def is_link_good(link): r = courlan.check_url(link,strict=True,language=LANGUAGE) if r is None: return None llink,lhostname = r #print(llink,lhostname) # hostname rules if not lhostname.endswith(DOMAIN): LOGGER.debug("bad hostname") return None if courlan.is_not_crawlable(llink): LOGGER.debug("not crawlable") return None return llink def get_link_doc(link,status="frontlink"): r = courlan.check_url(link) assert r is not None link,host = r domain = courlan.extract_domain(link) return {"url":link,"host":host,"domain":domain,"status":status,"created_at":datetime.utcnow()} def fetch_page(link): print("fetching:::::") print(link) final_link = link response = trafilatura.fetch_url(link,decode=False) time.sleep(2) html = None if response is not None : good = True if response.status != 200: good = False LOGGER.error('not a 200 response: %s for URL %s', response.status, url) elif response.data is None or len(response.data) < MINFILESIZE: LOGGER.error('too small/incorrect for URL %s', link) good = False # raise error instead? elif len(response.data) > MAXFILESIZE: good = False LOGGER.error('too large: length %s for URL %s', len(response.data), link) if good: html = trafilatura.utils.decode_response(response) final_link = response.url if html is not None: html, final_link = trafilatura.spider.refresh_detection(html, final_link) # is there a meta-refresh on the page? if final_link is None: # malformed or malicious content html = None return final_link,html def fetch_robot(base_url): try: rawrules = trafilatura.fetch_url("https://"+ base_url + "/robots.txt") #print(rawrules) rules = urllib.robotparser.RobotFileParser() rules.parse(rawrules.split("\n")) LOGGER.info('got robots') except Exception as exc: LOGGER.error('cannot read robots.txt: %s', exc) rules = None # exceptions happening here return rules def extract_pages(link_batch,responses): out = [] for original_link,(final_link,html) in zip(link_batch,responses): doc = None assert original_link is not None if html is not None: doc = trafilatura.bare_extraction(html,url=final_link,with_metadata=True,include_formatting=False,target_language=LANGUAGE,favor_precision=True) if doc is not None: if not "text" in doc or len(doc["text"]) < MINTEXTSIZE: # text too small doc = None out.append((original_link,final_link,html,doc)) return out def index_pages(db,hostname,extracted_pages): linkcol = db["links"] htmlcol = db["html"] contentcol = db["content"] checkcol = db["check"] links = [] for original_link,final_link,html,doc in extracted_pages: state = "good" link = original_link if original_link != final_link: linkcol.update_one({"url":original_link},{"$set":{"status":"redirect"}}) link = final_link if html is None: state = "html_error" elif doc is None: state = "content_error" if doc is not None: text = doc["text"] checksums,sizes = calculate_checksums(text) doc["text_size"] = len(text) doc["text_md5"] = hashlib.md5(text.encode("utf8")).hexdigest() doc["paragraph_checksums"] = checksums doc["paragraph_sizes"] = sizes goodsz = sum(sizes) # Not enough larger paragraphs if len(text) < TEXT_TRASH_SIZE or goodsz/len(text) < TEXT_TRASH_RATIO: state = "trash" end_sentence_marker = re.compile("\w[\.]") sentences = 0 for item in re.finditer(end_sentence_marker,text): t = item.group(0) if t[0].islower(): sentences += 1 doc["sentences"] = sentences # check copy if state == "good": copysz = len(text) - goodsz for chs,paragraph_size in zip(doc["paragraph_checksums"],doc["paragraph_sizes"]): # index paragraph checksums nd = checkcol.find_one({"_id":chs}) if nd is not None: copysz += paragraph_size if (copysz / len(text)) > TEXT_TRASH_RATIO: state = "copy" print(copysz) if state == "good": htdoc = get_link_doc(link,state) htdoc["html"] = html htdoc["html_size"] = len(html) htdoc["html_md5"]= hashlib.md5(html.encode("utf8")).hexdigest() # can be revisited - upsert del htdoc["url"] htmlcol.update_one({"url":link},{"$set":htdoc},upsert=True) doc.update(get_link_doc(link,"good")) # todo extract links print(doc) del doc["url"] contentcol.update_one({"url":link},{"$set":doc},upsert=True) for chs in doc["paragraph_checksums"]: try: checkcol.insert_one({"_id":chs}) except pymongo.errors.DuplicateKeyError as err: pass linkcol.update_one({"url":link},{"$set":{"status":state}}) def extract_links(link_batch,responses,hostname,rules,default_status="frontlink"): links = {} badrobot = 0 for original_link,(final_link,html) in zip(link_batch,responses): status = default_status external_links = courlan.extract_links(html,final_link,external_bool=True,language=LANGUAGE) for link in external_links: links[link] = "frontlink" internal_links = courlan.extract_links(html,final_link,external_bool=False,language=LANGUAGE) #print(extracted_links) for link in internal_links: if not is_robot_good(link,rules): badrobot += 1 continue status = str(default_status) #print(link,status) links[link] = status outlinks = [] badlink = 0 for link,status in links.items(): link = is_link_good(link) if link is None: badlink += 1 continue outlinks.append((link,status)) print(f"{len(links)} total links, {badrobot} badrobot {badlink} badlinks") return outlinks def index_links(db,extracted_links): linkcol=db["links"] for link,status in extracted_links: if not is_link_good(link): continue if status == "frontlink" or status == "backlink": doc = get_link_doc(link,status) try: linkcol.insert_one(doc) # dont overwrite except pymongo.errors.DuplicateKeyError as ex: pass else: print("updating " + link,status) linkcol.update_one({"url":link},{"$set":{"status":status,"updated_at":datetime.utcnow()}}) def get_link_features(link): a, urlpath = courlan.get_host_and_path(link) features = re.split("[/?&]",urlpath) #features = re.split("[/?-_=]",urlpath) res = [] for i,feature in enumerate(features): if len(feature) < 1: continue feature = re.sub("[0-9]","*",feature) res.append(str(i)+ "-" + feature) if len(res) < 2: return None res = res[:-1] print(res) return res class LinkClassifier: def __init__(self): self.goodcounter = collections.Counter() self.badcounter = collections.Counter() self.good_count = 0 self.bad_count = 0 self.alpha = 0.001 def train(self,links): for i,item in enumerate(links): link = item["url"] state = item["status"] cl = 0 if state == "good": cl = 1 print(cl,state,link) features = get_link_features(link) if features is None: continue lf = len(features) if state == "good": for feature in features: self.good_count += 1 self.goodcounter[feature] += 1 else: for feature in features: self.bad_count += 1 self.badcounter[feature] += 1 self.bdictsize = len(self.badcounter) self.gdictsize = len(self.goodcounter) def test(self,testset): # eval gg = 0 true_positive = 0 positive = 0 false_negative = 0 for item in testset: l = item["url"] cl = 0 if item["status"] == "good": cl = 1 pcp = self.classify(l) r = 0 if pcp > 0: r = 1 if cl == 1: if r == 1: true_positive += 1 positive += 1 if r == 1 and cl == 0: false_negative += 1 if r == cl: gg += 1 else: print("MISS",l,cl,pcp) print(len(testset)) print("Precision: {}, Recall: {}".format(true_positive/positive,true_positive/(true_positive+false_negative))) print("Accuracy:") acc = gg / len(testset) print(acc) return acc def classify(self,link): if self.good_count == 0 or self.bad_count == 0: return random.uniform(-0.1,0.1) features = get_link_features(link) res = 0 gp = math.log(self.good_count) - math.log(self.good_count + self.bad_count) bp = math.log(self.bad_count) - math.log(self.good_count + self.bad_count) if features is None: return math.exp(gp) - math.exp(bp) gcc = math.log(self.gdictsize * self.alpha + self.good_count) bcc = math.log(self.bdictsize * self.alpha + self.bad_count) goodprob = 0 badprob = 0 for feature in features: g = math.log((self.goodcounter[feature] + self.alpha)) - gcc goodprob += g b = math.log(self.badcounter[feature] + self.alpha) - bcc badprob += b print(feature,g,b) pa = math.exp(goodprob + gp) pb = math.exp(badprob + bp) return pa - pb #+ random.uniform(-0.001,0.001) def get_links(db,hostname,status,batch_size): linkcol = db["links"] res = linkcol.find({"host":hostname,"status":status},limit=batch_size) links = [] for item in res: links.append(item["url"]) print("Got {} {}".format(len(links),status)) return links def fetch_sitemap_links(start_link): out = [] navigation_links = trafilatura.sitemaps.sitemap_search(start_link,target_lang=LANGUAGE) for link in navigation_links: out.append((link,"frontlink")) print("Fetched {} sitemap links".format(len(out))) return out def fetch_front_links(start_link,rules): start_link,hostname = courlan.check_url(start_link) response = fetch_page(start_link) extracted_links = extract_links([start_link],[response],hostname,rules,"frontlink") print("Fetched {} frontlinks".format(len(extracted_links))) return extracted_links def link_summary(db,hostname): linkcol = db["links"] #res = linkcol.distinct("hostname",{"hostname":hostname}) res = linkcol.aggregate([ {"$match":{"host":hostname}}, {"$group":{"_id":"$status", "count":{"$count":{}}, } }, ]) badcount = 0 goodcount = 0 info = {} crawled_count = 0 bad_crawl_count = 0 for item in res: count = item["count"] st = item["_id"] print(st,count) if st == "good": goodcount += count if st != "frontlink" and st != "backlink": crawled_count += count if st != "good": bad_crawl_count += count info[st] = count info["crawled_count"] = crawled_count info["bad_crawl_count"] = bad_crawl_count baclink_cout = 0 if "backlink" in info: backlink_count = info["backlink"] good_prob= 0 if crawled_count > 0: good_prob = goodcount / crawled_count info["good_prob"] = good_prob print(">>>Domain Content") contentcol = db["content"] res = contentcol.aggregate([ {"$match":{"host":hostname}}, #{"$project": {"textsum":{"$sum":"$text_size"}}} {"$group":{"_id":None, "text_size_sum":{"$sum":"$text_size"}, } }, ]) text_size = 0 for item in res: text_size = item["text_size_sum"] good_document_characters = 0 fetch_average_characters = 0 if goodcount > 0: good_document_characters = text_size / goodcount fetch_average_characters = text_size / crawled_count info["total_good_characters"] = text_size info["average_good_characters"] = good_document_characters info["average_fetch_characters"] = fetch_average_characters domaincol = db["domain"] domaincol.update_one({"host":hostname},{"$set":info},upsert=True) res = domaincol.find_one({"host":hostname}) print(res) def sample_links(db,hostname,status,batch_size): print("Getting backlinks") linkcol = db["links"] res = linkcol.find({"host":hostname,"status": {"$not":{"$in":["frontlink","backlink"]}}}) cl = LinkClassifier() crawled_links = list(res) crawled_count = len(crawled_links) prediction_accuracy = 0 if crawled_count > 200: # train on crawled links trainset,testset = split_train(crawled_links) cl.train(trainset) prediction_accuracy = cl.test(testset) sample_set_size = 10000 res = linkcol.find({"host":hostname,"status": status},limit = sample_set_size) sample_links = [] predicted_good = 0 for item in res: for item in res: cll = cl.classify(item["url"]) #cll += random.uniform(-0.1,0.1) sample_links.append((item["url"],cll)) if cll > 0: predicted_good += 1 # TODO frontlinks are not unique! sample_links.sort(key=lambda x: x[1],reverse=True) predicted_good_prob = 0 if len(sample_links) > 0: predicted_good_prob = predicted_good / len(sample_links) domaincol = db["domain"] info = { "predicted_good_prob":predicted_good_prob, "prediction_accuracy": prediction_accuracy, "crawled_count": crawled_count, } print(info) domaincol.update_one({"host":hostname},{"$set":info}) links = [l[0] for l in sample_links[0:batch_size]] return links def domain_summary(db,hostname): linkcol = db["links"] #res = linkcol.distinct("hostname",{"hostname":hostname}) # count links res = linkcol.aggregate([ {"$group":{"_id":"$hostname","text_size_sum":{"$sum":"$text_size"}}}, ]) for item in res: print(item) @click.group() def cli(): pass @cli.command() def createdb(): myclient = pymongo.MongoClient(CONNECTION) db=myclient[DBNAME] linkcol = db["links"] linkcol.create_index("url",unique=True) linkcol.create_index("host") contentcol = db["content"] contentcol.create_index("url") contentcol.create_index("text_md5",unique=True) #contentcol.create_index({"paragraph_checksums":1}) contentcol.create_index("host") htmlcol = db["html"] htmlcol.create_index("url") htmlcol.create_index("html_md5",unique=True) domaincol = db["domains"] domaincol.create_index("host",unique=True) domaincol.create_index("average_fetch_characters",unique=True) @cli.command() @click.argument("link") def parseurl(link): link,hostname = courlan.check_url(link) rawrules = trafilatura.fetch_url("https://"+ hostname + "/robots.txt") print(rawrules) rules = urllib.robotparser.RobotFileParser() rules.parse(rawrules.split("\n")) print(rules.can_fetch("*",link)) print(rules.site_maps()) print(rules.crawl_delay("*")) html = trafilatura.fetch_url(link,decode=True) doc = trafilatura.bare_extraction(html) import pprint pprint.pprint(doc) @cli.command() @click.argument("link") def externaldomains(link): html = trafilatura.fetch_url(link,decode=True) external_links = courlan.extract_links(html,link,external_bool=True,language=LANGUAGE) domains = set() for l in external_links: r = courlan.check_url(l) if r is None: pass link,domain = r domains.add(domain) for d in domains: print(d) @cli.command() @click.argument("start_link") def classify(start_link): myclient = pymongo.MongoClient(CONNECTION) db=myclient[DBNAME] start_link,hostname = courlan.check_url(start_link) cl = LinkClassifier() linkcol = db["links"] res = linkcol.find({"host":hostname,"status": {"$not":{"$in":["frontlink","backlink"]}}}) trainset, testset = split_train(res) cl.train(trainset) cl.test(testset) @cli.command() @click.argument("start_link") def visit(start_link): myclient = pymongo.MongoClient(CONNECTION) db=myclient[DBNAME] start_link,hostname = courlan.check_url(start_link) batch_size = BATCHSIZE rules = fetch_robot(hostname) # renew front links front_links = fetch_front_links(start_link,rules) index_links(db,front_links) # start crawling # frontlinks first links = sample_links(db,hostname,"frontlink",batch_size) links.insert(0,start_link) # then backlinks if len(links) < batch_size: back_links = sample_links(db,hostname,"backlink",batch_size - len(links)) links += back_links # index results print("Processing links") responses = [] for link in links: responses.append(fetch_page(link)) extracted_pages = extract_pages(links,responses) extracted_links = extract_links(links,responses,hostname,rules,"backlink") index_links(db,extracted_links) index_pages(db,hostname,extracted_pages) link_summary(db,hostname) if __name__ == "__main__": cli()