#!/usr/bin/env python2 # -*-coding:UTF-8 -* """ The ZMQ_Sub_Curve Module ============================ This module is consuming the Redis-list created by the ZMQ_PubSub_Tokenize_Q Module. Each words files created under /files/ are representing categories. This modules take these files and compare them to the stream of data given by the ZMQ_PubSub_Tokenize_Q Module. When a word from a paste match one or more of these words file, the filename of the paste is published/forwarded to the next modules. Each category (each files) are representing a dynamic channel. This mean that if you create 1000 files under /files/ you'll have 1000 channels where every time there is a matching word to a category, the paste containing this word will be pushed to this specific channel. ..note:: The channel will have the name of the file created. Implementing modules can start here, create your own category file, and then create your own module to treat the specific paste matching this category. ..note:: Module ZMQ_Something_Q and ZMQ_Something are closely bound, always put the same Subscriber name in both of them. Requirements ------------ *Need running Redis instances. (Redis) *Categories files of words in /files/ need to be created *Need the ZMQ_PubSub_Tokenize_Q Module running to be able to work properly. """ import redis, argparse, zmq, ConfigParser, time from packages import Paste as P from packages import ZMQ_PubSub from pubsublogger import publisher from packages import lib_words configfile = './packages/config.cfg' def main(): """Main Function""" # CONFIG # cfg = ConfigParser.ConfigParser() cfg.read(configfile) # SCRIPT PARSER # parser = argparse.ArgumentParser( description = '''This script is a part of the Analysis Information Leak framework.''', epilog = '''''') parser.add_argument('-l', type = str, default = "../files/list_categ_files", help = 'Path to the list_categ_files (../files/list_categ_files)', action = 'store') args = parser.parse_args() # REDIS # r_serv = redis.StrictRedis( host = cfg.get("Redis_Queues", "host"), port = cfg.getint("Redis_Queues", "port"), db = cfg.getint("Redis_Queues", "db")) r_serv1 = redis.StrictRedis( host = cfg.get("Redis_Level_DB", "host"), port = cfg.get("Redis_Level_DB", "port"), db = 0) # LOGGING # publisher.channel = "Script" # ZMQ # channel = cfg.get("PubSub_Words", "channel_0") subscriber_name = "curve" subscriber_config_section = "PubSub_Words" Sub = ZMQ_PubSub.ZMQSub(configfile, subscriber_config_section, channel, subscriber_name) # FUNCTIONS # publisher.info("Script Curve subscribed to channel {0}".format(cfg.get("PubSub_Words", "channel_0"))) paste_words = [] message = Sub.get_msg_from_queue(r_serv) prec_filename = None while True: if message != None: channel, filename, word, score = message.split() if prec_filename == None or filename != prec_filename: PST = P.Paste(filename) lib_words.create_curve_with_word_file(r_serv1, "/home/user/AIL/var/www/static/csv/wordstrendingdata", "/home/user/AIL/files/wordfile", int(PST.p_date.year), int(PST.p_date.month)) prec_filename = filename prev_score = r_serv1.hget(word.lower(), PST.p_date) print prev_score if prev_score != None: r_serv1.hset(word.lower(), PST.p_date, int(prev_score) + int(score)) else: r_serv1.hset(word.lower(), PST.p_date, score) #r_serv.expire(word,86400) #1day else: if r_serv.sismember("SHUTDOWN_FLAGS", "Curve"): r_serv.srem("SHUTDOWN_FLAGS", "Curve") print "Shutdown Flag Up: Terminating" publisher.warning("Shutdown Flag Up: Terminating.") break publisher.debug("Script Curve is Idling") print "sleepin" time.sleep(1) message = Sub.get_msg_from_queue(r_serv) if __name__ == "__main__": main()