import json import re import time from .common import InfoExtractor from ..utils import month_by_name class NDTVIE(InfoExtractor): _VALID_URL = r'^https?://(?:www\.)?ndtv\.com/video/player/[^/]*/[^/]*/(?P[a-z0-9]+)' _TEST = { u"url": u"http://www.ndtv.com/video/player/news/ndtv-exclusive-don-t-need-character-certificate-from-rahul-gandhi-says-arvind-kejriwal/300710", u"file": u"300710.mp4", u"md5": u"39f992dbe5fb531c395d8bbedb1e5e88", u"info_dict": { u"title": u"NDTV exclusive: Don't need character certificate from Rahul Gandhi, says Arvind Kejriwal", u"description": u"In an exclusive interview to NDTV, Aam Aadmi Party's Arvind Kejriwal says it makes no difference to him that Rahul Gandhi said the Congress needs to learn from his party.", u"upload_date": u"20131208", u"duration": 1327, u"thumbnail": u"http://i.ndtvimg.com/video/images/vod/medium/2013-12/big_300710_1386518307.jpg", }, } def _real_extract(self, url): mobj = re.match(self._VALID_URL, url) video_id = mobj.group('id') webpage = self._download_webpage(url, video_id) filename = self._search_regex( r"__filename='([^']+)'", webpage, u'video filename') video_url = (u'http://bitcast-b.bitgravity.com/ndtvod/23372/ndtv/%s' % filename) duration_str = filename = self._search_regex( r"__duration='([^']+)'", webpage, u'duration', fatal=False) duration = None if duration_str is None else int(duration_str) date_m = re.search(r'''(?x) \s* Published\s+On:\s* (?P[A-Za-z]+)\s+(?P[0-9]+),\s*(?P[0-9]+) ''', webpage) upload_date = None assert date_m if date_m is not None: month = month_by_name(date_m.group('monthname')) if month is not None: upload_date = '%s%02d%02d' % ( date_m.group('year'), month, int(date_m.group('day'))) description = self._og_search_description(webpage) READ_MORE = u' (Read more)' if description.endswith(READ_MORE): description = description[:-len(READ_MORE)] return { 'id': video_id, 'url': video_url, 'title': self._og_search_title(webpage), 'description': description, 'thumbnail': self._og_search_thumbnail(webpage), 'duration': duration, 'upload_date': upload_date, }